Price Control and Access to Drugs:

The Case of India’s Malarial Market

Debi Prasad Mohapatra∗†

Chirantan Chatterjee‡

Abstract

This article investigates the effects of drug price-control policy on access to drugs and consumer
welfare. We focus on regional markets for malarial drugs in India. Governments in developing and
underdeveloped countries implement price controls to make drugs affordable and improve consumer
welfare. However, gains from lower prices due to price control have little meaning if, firms stop
selling those drugs in response to price control, and drugs are not available for purchase in local
markets. In order to characterize the trade-off between availability and affordability on consumer
welfare, we estimate a structural model of Indian malarial market where firms make endogenous
product offering decisions across regions. We use a unique dataset that features detailed region level
information. Our estimates reveal that even in a poor region, lower prices lead to lower consumer
welfare, as costs of making a drug available in a regional market are high enough to induce exit of
products in response to lower prices. We characterize the optimal price control level that balances
the trade-off and maximizes consumer welfare.
Key Words: Price Control, Malaria, Access to Drugs, Regional Markets, Endogenous Product
Choices, Inequality Estimation

∗
205B Stockbridge Hall, University of Massachusetts, Amherst, dmohapatra@umass.edu
†
Mohapatra is grateful to his dissertation committee members, Panle Jia Barwick, David Easley, Matt
Backus, Larry Blume, and Francesca Molinari for invaluable help, advice and encouragement. He sincerely
thanks Nikhil Agrawal, Levon Barseghyan, Chris Conlon, Michael Dickstein, Amit Gandhi, Hiroaki Kaido,
Ryan Lampe, Priya Mukherjee, Charles Murry, Joris Pinkse, Nancy Rose, Steven Stern, Joerg Stoye, Thomas
Wollmann, Mo Xiao, Yang Zhang, and seminar participants in Cornell, MIT, Rochester Simon school, UMass
Amherst, Microsoft Research, SUNY Albany, Northeastern, SUNY Stony Brook, U of Washington, and
ISB Hyderabad for comments and helpful suggestions. Chatterjee acknowledges support from the All India
Organization of Chemists and Druggists for sharing the data and from the Max Institute and Bharti Institute
Research Fellowships in Healthcare and Public Policy at the Indian School of Business.
‡
Economics and Public Policy, Indian School of Business, India
1 Introduction
Access to drugs is a contentious issue in the context of developing and underdeveloped coun-
tries, where limited access has excluded many patients from the benefits of pharmaceutical
products. Much of the debate about access has focused on the trade-offs between welfare
gains from innovation incentives (for example, patents) and resulting welfare loss due to
higher price (Moser (2013), Chaudhuri, Goldberg, and Jia (2006), Kyle and Qian (2014),
Duggan, Garthwaite, and Goyal (2016)). Due to limited insurance coverage and poorly
funded public health systems in low-income countries, high prices may act as a barrier
for large segments of the population against affordable access to developed drug therapies.
Hence, economists have recommended maintaining low drug prices while providing incen-
tives for innovation through public interventions or through external efforts (Kremer (1998,
2002)). To promote affordable access, various low-income countries have adopted price con-
trol policies to lower prices and potentially improve consumer welfare.
This debate, however, misses a critical element: even when there are no incentive prob-
lems in R & D investments, gains from lower prices due to price control would not benefit
consumers if, firms stop selling those drugs in response to price control, and drugs are not
available to consumers for purchase in local markets. Hence, while improved affordability due
to price control may increase access and consumer welfare, lack of availability may lead to
decreased access and lower consumer welfare. As such, the effect of price control on access to
drugs and consumer welfare is a question that calls for empirical investigation. Even though
price control policies are widely adopted in various developing countries including China,
India, Indonesia, Philippines, and Thailand among others, rigorous empirical evidence on
welfare implications of such policies is mostly missing in the current literature. In this paper,
we aim to partially fill this gap in the literature by examining the effects of a price control
policy in India while accounting for firms’ incentives to adjust their product portfolios in the
market.
Our analysis is done in the context of regional markets for malarial drugs in the Indian
pharmaceutical industry. We use a new and unique dataset that records sales and prices of
drugs sold across 23 different regions in India from 2007 to 2013. Since we observe product
offering as well as price setting1 decisions by the firms, we can ask: “What would have hap-
pened to product offering and consumer welfare, had the government imposed price control
during this period? What is the optimal level of price control that maximizes consumer
welfare?” To examine these questions, we estimate a model of supply and demand where
1
During our sample period, only 3 out of 11 molecules were under price control. For other molecules,
firms were free to set prices.

1
both the set of products offered in the regional markets by a firm and their prices are en-
dogenously determined. Regional disaggregation of our dataset provides us with a unique
advantage over most of the existing studies on the Indian pharmaceutical industry, since we
use the significant variation across different regions to estimate our model. We model a two-
stage game played by the drug manufacturers: in the first stage firms face a discrete menu
of molecules and simultaneously choose which set of products to offer in each region. In the
second stage, the chosen products are sold to the consumers in a simultaneous-price-setting
game. While consumer heterogeneity and profitability provide incentives to firms to offer
multiple products in a region, offering each of these products results in fixed costs. In our
empirical investigation, we therefore need to estimate both expected variable profit and fixed
costs.
To learn expected variable profit, we estimate a random coefficient logit model to recover
the distribution of consumer preferences following Berry, Levinsohn, and Pakes (1995, BLP,
hereafter) and Nevo (2001). We recover marginal costs for drug production using equilibrium
first-order conditions resulting from firms’ profit maximization. To learn total fixed cost, we
use a revealed preference argument commonly used in empirical entry literature, specifically,
the fact that, a firm offers a product only if its variable profit exceeds the corresponding fixed
cost. Naturally, this condition yields a selected sample of offered products, and we address
the associated endogeneity problem (“selection problem” as discussed in Pakes, Porter, Ho,
and Ishii (2015, PPHI, hereafter)) by proposing a novel instrument that follows the monotone
instrumental variables approach developed in Manski and Pepper (2000, 2009).
Our estimates reveal the presence of heterogeneity across different regions in terms of
demand characteristics such as elasticities and willingness to pay. Consumers in high-income
regions are on average less price sensitive compared to consumers in low-income regions. In
addition, our estimates indicate that fixed costs vary across regions and across firms, and
more importantly, on average, are significant proportions of variable profits. This suggests
that fixed costs may play an important role in product offering decisions of the firms. We
then perform counterfactual analyses to evaluate the impact of price control on access to
drugs and consumer welfare. We fix the price-caps by setting a specific markup level above
estimated marginal cost. For the products under price control, we allow firms to set the
prices while not exceeding those respective price-caps. We simulate firm’s product entry
decisions at the specific markup, and compute corresponding consumer welfare. We repeat
this exercise while varying the markup levels from 1%, 2%, . . . , up to 30% for Odisha, one
of the poorest regions in India that features highest number of malarial incidence.
Our analysis reveals that, fixed costs are large enough to induce exit of products from
the market in response to a stringent price control policy. As we decrease the margins and

2
reduce prices, firms also withdraw their products. Since consumer welfare depends both
on prices as well as product availability, with varying markup levels, the total consumer
welfare follows an inverted-U shaped pattern. In our counterfactual exercise, starting from
1%, as we increase markup level for price control, more products are made available for
purchase in the market. Hence, even in a poor region, consumer welfare increases with
higher prices up to 16% markup. Our results show that, further increase in prices, allowing
markup beyond 16%, leads to lowering of consumer welfare. This is because, the welfare
loss due to higher prices dominates the welfare gain from additional product availability,
and overall consumer welfare goes down.2 Our framework makes it possible to recover these
findings and characterize optimal price control level, since it combines a study of endogenous
product decisions and prices with a detailed analysis of cost and demand.
Our study focuses on ‘already developed drugs’, and unintended consequences of lowering
those prices through price control policies. In contrast to a growing literature on innovation
in pharmaceuticals, we demonstrate that, even when there are no inefficiencies from R &
D incentives, concerns from distributional inefficiencies may have major consequences for
consumer welfare. Our choice of studying malaria fits into this focus because of several
reasons. First, malaria is an important neglected disease with a very high disease burden3 in
India (see Kumar, Valecha, Jain, and Dash (2007)). Second, it is widely reported that the
parasite causing malaria is mostly resistant to older and relatively cheaper drugs. New and
more effective drugs are however relatively highly priced. Moreover, efficacy and suitability
of the drugs vary across consumers depending on age, gender, and health conditions. Hence,
to fight this disease, multiple antimalarial therapies are widely recommended. That is why,
we focus on the availability and affordability of those new malarial drugs which carry high
welfare significance. Third, our estimation framework is especially suited for malaria, as all
these drugs are already-developed generic drugs, and making them available as well as fixing
respective prices are the only choices made by the firms.4 Hence, the firms’ decisions on
availability and pricing affect consumers’ access to these developed drugs. Finally, number
of malarial patients depends on prevalence of mosquitoes which is in turn varies with rainfall
and other seasonal factors. Hence, across different regions, demand for malarial drugs varies
significantly over the seasons in a year. In our sample, under free entry condition and without
price control, we observe a stable set of firms actively making product offering decisions in
2
In contrast, the government mandated price control rule (under ‘Drug Price Control Policy 1995’) fixes
the markup to manufacturer at around 8-10%.
3
Disease burden is a measure that combines years of life lost due to premature mortality and years of life
lost due to time lived in states of less than full health (WHO).
4
Research and development of malarial drugs is primarily funded by several international organizations
(like Medicines for Malaria Venture (MMV), Drugs for Neglected Diseases initiative (DNDi), Bill and Melinda
Gates foundation). Once the drugs are developed, the marketing rights are provided to different firms.

3
response to these changing demand conditions. This variation helps us to estimate the profit
function of the firms under reasonable assumptions.5

1.1 Methodology and Related Literature

The model in this article has the following key features: (i) It allows firms to make multiple
discrete-product choices, so that both the number and identity of offered product are treated
as endogenous. (ii) It incorporates a detailed model of differentiated-product cost and de-
mand system. (iii) It allows for flexible firm-region-product specific fixed cost, and proposes
a new way to deal with the resulting endogeneity bias issues. The estimation methods used
in this article belong to a growing literature on entry, exit and endogenous product choices
(starting from Mazzeo (2002), and Seim (2006), Draganska, Mazzeo, and Seim (2009)). More
recently, Fan (2013) demonstrates that taking into account endogenous product quality de-
cisions following a merger leads to substantial differences in estimated effects of mergers.
Similarly, in the home PC market, allowing for consumer heterogeneity and endogenous
product offering decisions by PC manufacturers, Eizenberg (2014) demonstrates that the
welfare effects of innovation has heterogeneous implications on consumers. In case of CPU
market, Nosko (2011) studies product line decisions by AMD and Intel taking into account
asymmetric technological capabilities and varying competitive environments. In the context
of $85 billion automotive industry bailout, Wollmann (2017) shows that ignoring endogenous
product choice decisions by the firms following a merger may lead to substantially different
ex-post outcomes. Similarly, Fan and Yang (2016) studies the effects of changing competi-
tion on number and composition of product offering in the US smart-phone market.6 Our
paper adds to this literature by studying the effects of endogenous product choice decisions
in the context of a widely implemented regulation in a developing country like India.
In our model of product entry, we assume complete information and employ Subgame
Perfect Nash Equilibrium as solution concept. As it is well established in the entry liter-
ature (e.g. Tamer (2003)), uniqueness of equilibrium is not guaranteed, leading to partial
identification of fixed costs (Andrews, Berry, and Jia (2004), Ciliberto and Tamer (2009),
Beresteanu, Molchanov, and Molinari (2011)). We exploit the necessary equilibrium condi-
tions to place bounds on partially identified parameters following PPHI, and Ho and Pakes
(2014). Most of the applications in entry literature employ a reduced-form profit function,
whereas we derive the profit function from micro-foundations with a detailed model of cost
and demand. We use techniques recently developed in Kaido, Molinari, and Stoye (2016), to
obtain element-wise confidence intervals on the fixed cost parameters that are asymptotically
5
We describe the details as well as evidences in section 4.
6
See Crawford (2012) for an excellent recent survey.

4
uniformly valid.
A sample selection problem arises in the entry model, as firms are explicitly assumed
to have selected the set of products observed in the data. Such selection problems are ex-
tensively discussed in PPHI, Eizenberg (2014), and Dickstein and Morales (2013) among
others. Following Manski and Pepper (2000), we propose a monotone instrumental vari-
able approach to address this problem. Specifically, we assume mean monotonicity of the
selected error terms conditional on observed characteristics and use those characteristics as
instruments to solve this problem. The intuition behind our proposal is straightforward.
Firms sell not only malarial drugs, but also drugs for close therapeutic categories like cold,
fever and antibiotics. If a firm enjoys a high revenue share in selling these closely related
products compared to its rival, then it is reasonable to assume that, the firm also enjoys
better facilities (like better arrangement of electricity for storage of drugs, better relation-
ship with distributors as well as doctors, better network of medical representatives) while
offering malarial drugs. Our mean monotonicity assumption then implies that, conditional
on higher presence in close therapeutic categories, and controlling for observable character-
istics, compared to a rival, a firm enjoys more favorable fixed cost shock in expectation. We
can then use those monotonicity conditions to construct moments for our estimation.
Finally, this article contributes to a growing literature on the effect of price control and
government policy-related topics in the context of developing countries and their medicine
markets (e.g. see Goldberg (2010), Duggan, Garthwaite, and Goyal (2016), Kessler (2004),
Morton (1999), Dubois, De Mouzon, Scott-Morton, and Seabright (2015), Lanjouw (2005,
1998), Chatterjee, Kubo, and Pingali (2015), Dubois and Saethre (2018)). Through theo-
retical models (for example, Bond and Saggi (2014), Filson (2012)), and empirical studies
relying on cross-country observations (for example, Kyle and Qian (2014), Kyle and McGa-
han (2012), Danzon, Mulcahy, and Towse (2015), and Cockburn, Lanjouw, and Schankerman
(2016)), these papers clearly demonstrate the inherent trade-offs faced by government poli-
cies, in the context of India as well as various other countries. For example, Kyle (2007)
finds that the use of price controls has a statistically and quantitatively important effect
on the extent and timing of the launch of new drugs. Similarly, Danzon, Wang, and Wang
(2005) analyzes the effect of pharmaceutical price regulation on delays in new drug launches
across 25 different markets. We add to this literature by providing a detailed structural
analysis of consumer heterogeneity, and firm behavior, and by studying the effects of poli-
cies to maximize consumer welfare using micro-data. In closely related work, Chaudhuri,
Goldberg, and Jia (2006) study the Quinolones antibiotic segment in India, and investigate
the welfare implications of patent policy while allowing firms to adjust prices. Dutta (2011)
also addresses welfare implications of patent policy by allowing firms to respond to policy

5
changes, while treating all firms as homogeneous single product units. In contrast, we use re-
gionally disaggregated data, allow full heterogeneity across firms and regions, and study the
welfare implications of price control while allowing firms to readjust both product offerings
and prices in response to the policy change.
The rest of the article is organized as follows: Section 2 describes industry and the data.
Section 3 presents the model. Section 4 discusses identification and estimation. Section
5 discusses the results from our estimation. Section 6 reports results from counterfactual
analysis. Section 7 concludes.

2 Industry Background and Data

In this section, we briefly describe malaria, its prevalence in India, and set up of price control
policy. We then provide details of our data and present descriptive statistics.

2.1 Malaria in India

Malaria is a mosquito-borne infectious disease caused by parasitic protozoans (a type of sin-
gle cell micro-organism) of the plasmodium type. According to the 2015 WHO report titled
‘Achieving the Malaria Millennium Development Goals Target’7 , in terms of deaths due to
malaria, India ranks third in the world after Nigeria and Democratic Republic of Congo.
Taking into account the lost earnings due to bad health, as well as treatment costs, apart
from mortality, malaria imposes an economic burden of US$1940 million in India (estimates
by Gupta, Chowdhury, et al. (2014)). These numbers are baffling, since drugs for effective
malarial treatment exist and are sold by different firms in India. The drugs to treat malaria
are also not under patent protection,8 as non-profit organizations (NPOs) like Drugs for
Neglected Diseases initiative (DNDi) and Medicines for Malaria Venture (MMV) generally
raise funds from different foundations (like Bill and Melinda Gates foundation) and work
with a commercial partner to conduct Research and Development. These NPOs retain the
intellectual property right and provide marketing rights to firms to ensure effective distribu-
tion of drugs across different countries.9 Since with timely detection and with appropriate
drugs, malaria is curable, it seems that even if effective drugs are developed, they may not
reach the final consumers due to various inefficiencies.
[Table 1 about here]
7
Organization, UNICEF, et al. (2015)
8
In our sample, for each molecule, multiple firms produce generic products. Note that, under patent
protection, only the patent holder has the exclusive marketing right.
9
See the intellectual property policy of MMV for more details.

6
 Although malaria can happen due to multiple strains of the parasite, according to the
U.S. Center for Disease Control, more than 90% of the malaria cases in India is caused due
to P. vivax and P. falciparum. Table 1 lists the 11 molecules we observe in our data, and
summarizes the average prices and quantities sold for the molecules in our sample. Tradi-
tionally, malaria in India has been treated with molecules like Chloroquine, Pyrimethamine-
Sulphadoxine and Hydroxychloroquine. As table 1 shows, these molecules are widely sold in
the India and are available at low prices. We refer to these molecules as old molecules, as
these molecules are being sold in India since 1970s. Drug price control policy 1995 also puts
these molecules under price control. However, malarial parasites (P. falciparum and P. vivax)
have developed resistance to these molecules leading to increasing cases of treatment failure
(see Antony and Parija (2016) for a recent study on malarial resistance). Recently, more
effective artemisinin-based combination therapies (ACTs) are developed to treat malaria.
In our sample, these new and effective molecules include Arteether-Artemotil, Artesunate,
Artemether-Lumefantrine, Artesunate-Mefloquine, and Arteether-Lumefantrine. The effi-
cacy and suitability of these molecules vary across patients depending on age, gender, as
well as physical conditions. Hence, consumers benefit when multiple new molecules are made
available in local stores at affordable prices. Additionally, use of combination of antimalarial
therapies, particularly with atreminisin derivatives can bring added advantage to consumers
by delaying resistance by malarial parasite (see Petersen, Eastman, and Lanzer (2011), Boni,
Smith, and Laxminarayan (2008) for a discussion on multiple treatments). Given the welfare
significance, we focus on availability and affordability of these new molecules across different
regions in India especially in the context of price control policy.

2.2 Price Control in India

Price control is one of the key instruments used by government of India with a goal to
ensure affordability of drugs. According to the National Pharmaceutical Pricing Authority
of India, 652 drugs, covering close to 35% of the Indian pharmaceutical market was under
price control in 2015. Especially after ratification of patent protection in India through
adoption of TRIPs,10 price control has been used extensively with a goal to improve access
to drugs.
Control of medicine prices in India is guided by Drug Price Control Order (DPCO) 1995
and Drug Price Control Order 2013. According to DPCO 1995, the retail price of a medicine
is calculated by the government in accordance with the following formula:
10
TRIPs stands for Trade Related Intellectual Property Rights. Under WTO agreement, India adopted
TRIPs to ensure patent protection in different industries including pharmaceuticals.

7
 Retail Price = (Cost Estimate) × (1 + margin/100)

where cost estimate would include material cost, packaging cost, and other manufacturing
costs including taxes. Typically, the profit margin to the manufacturer is set at around 8%.11
DPCO 1995 mandates each drug manufacturer in India to display its price on the pack of
the drug (also referred to as maximum retail price) and this is the price consumers pay while
purchasing the drug. It is important to point out here that, for a given drug under price
control (say 150mg tablet), a single price is fixed for all regions in India.12 Since during
our sample period prices were controlled under DPCO 1995, in our analysis, we use the cost
based price control rule from 1995, and investigate the trade-off between lower prices and
incentives of the firms to make the drug available in local markets while varying margin
values. It is worth noting here that, bringing further amendments to drug price control
order is being actively debated both by government bodies as well as in the associations of
pharmaceutical companies.13 In the next section we discuss our data sources and describe
key elements of our data.

2.3 Data: Source and Description

We obtained our primary data from the AIOCD’s (All India Organization of Chemists and
Druggists) subsidiary marketing research company AIOCD Awacs Pvt. Ltd. Additionally,
we collect consumptions survey data from NSSO (National Sample Survey Organization)
India, data on malaria patients from Indian National Vector Borne Disease Control website,
and data on rainfall from Indiastat, a data source provided by government of India. The
AIOCD data are arguably more accurate relative to the IMS data, an alternative private
source, as the AIOCD has better coverage and compliance in the collection of sales and
price data.14 Additionally, this dataset is widely used by financial analysts as well as by
Competition Commission of India for examination of its anti-trust cases. Our data records
prices and sales of all the medicines sold in India between March 2007 and September 2013.
An important aspect of our data is the disaggregation at the regional level among 23 geo-
graphic markets carved out by the AIOCD. It is worth pointing out that most of the current
studies on Indian pharmaceutical Industry including Chaudhuri, Goldberg, and Jia (2006)
and Dutta (2011) use price and sales data from IMS which divides India into four broad
11
This formula as well as similar profit margin was also used by Dutta (2011) in her calculations.
12
In 2013, price control policy in India was revised. In 2015 and in 2016, some further amendments were
also made in the price control policy.
13
For example see articles in The Times of India. The Mint.
14
This dataset has also been used in Mehta, Farooqui, and Selvaraj (2016), and Bhaskarabhatla, Chatter-
jee, and Karreman (2016).

8
divisions.15 The finer geographic disaggregation of AIOCD data provides us with an unique
advantage over IMS data, as we use the variations across different regions in India to study
the incentives of the firms while making product offering decision across different regional
markets in India.
We define a region-quarter combination as a market. In a given market, pharmaceutical
products are available in multiple presentations, that is, combination of dosage forms (e.g.
tablets and injections), strength ( e.g. 100 milligrams, 500 milligrams) and packet sizes (e.g.
50 tablet bottle, 100 tablet bottle). The various presentations in which a product is available
are often referred to as stock-keeping-units (SKUs). We focus our study on 11 molecules
that are used to cure malaria. Our data contain information on firm-level sales as well as
maximum retail prices (MRP) of these molecules at the stock keeping unit (SKU) level.
Since our analysis is focused on product offering decisions of the firms as well as consumer
preferences for the products, in our analysis, we aggregate SKUs for tablets and for injections
for a given firm for a specific molecule and define it as a “product”. Hence, our product
is defined at the level of firm-molecule-tablet and firm-molecule-injection.16 The details of
data preparation are discussed in appendix A.1.
We now discuss variation of prices across firms as well as across molecules in our sample.
In our data, for a given SKU, a firm charges identical price across all regions in India at a
given point in time, even when the molecule is not under price control. Given that regions
in India vary in demand characteristics (like number of malarial patients, income levels),
a firm’s decision to charge identical price across all regions may seem counter-intuitive.
However, firms charge same price across regions typically to avoid menu costs.17 In our
sample, prices vary over time as well as across firms. For a SKU for a given molecule, with
identical pack size and strength information, prices tend to differ significantly across firms.18

[Table 2 about here]

Table 2 documents average prices, standard deviation of prices, as well as minimum and
maximum prices for different molecules in our sample. Out of these 11 molecules, three
15
IMS data divides India into North, South, East and West.
16
For example, if IPCA pharmaceuticals produces 10 pack and 20 pack tablets for chloroquine, we combine
these SKUs for tablets to define a product as IPCA-Chloroquine-Tablet.
17
Since firms have to print the maximum retail price on the pack, printing different prices for different
regions and making sure transportation of the products with correct price to the relevant regions may involve
significant menu costs. We came to know about this during our interactions with the industry experts. It is
worth pointing out that, DellaVigna and Gentzkow (2017) document very similar patterns of uniform pricing
for food, drugstore, and mass merchandise chains in the US as well.
18
For example, the price of 60 mg injection of Artesunate varies between 73 Indian Rupees and 250 Indian
rupees depending on the firm producing the SKU. We manually conducted external validation of prices at
SKU level by consulting price data from websites CIMS India, and https://www.1mg.com/.

9
molecules, Pyrimethamine-Sulphadoxine, Chloroquine, and Hydroxychloroquine are under
price control since 1995 (under Drug Price Control Order, 1995). We highlight two impor-
tant points from this table. First, the average prices of new molecules including Arteether-
Artemotil, Artesunate, Artemether-Lumefantrine, Artesunate-Mefloquine, and Arteether-
Lumefantrine are around 15 to 20 times higher compared to average prices of old molecules
including Chloroquine and Pyrimethamine-Sulphadoxine. This suggests, without any price
control, consumers, especially with low income, might prefer old drugs over new drugs even
when the old drugs are not as effective. Second, among the new molecules, prices vary signif-
icantly across firms. For example, in case of Arteether-Artemotil, a popular new molecule,
price for a complete treatment varies between 38 Indian Rupees and 390 Indian Rupees. The
price variation arises since, unlike developed country markets, in India for a given molecule,
products sold by different firms are treated as differentiated products. Although the same
generic molecule is produced by different firms, products are typically grouped into branded
generic and non-branded generic products. These products vary in terms of brand name,
actual as well as perceived quality, packaging, and also in terms of number of available va-
rieties. These observations are also documented in other studies including Shrank, Cox,
Fischer, Mehta, and Choudhry (2009), and Basak and Sathyanarayana (2012). We use this
motivation to model the market as a differentiated product market while estimation of de-
mand. We must recognize here that a part of the price variation is artificially created due
to grouping of SKUs across solid and liquid forms. As an example, if price of 20 pack tablet
is not double the price of 10 pack tablet,19 then per tablet price may vary while combining
different types of presentations.20
Finally, we briefly discuss the drug distribution system in India. Drug manufacturers
usually operate in each region separately through regional warehouses (in industry termi-
nology they are known as carry forward agents). Since crossing a state border and serving
a separate state incurs additional taxes, a regional warehouse in a state usually serves the
distributors inside a state. This fits well into our analysis, as firms take product offering
decisions at the region level, and our data is also disaggregated across regions. Usually a firm
operates through 1-3 warehouses in each state in India. Cost of the distribution from man-
19
Typically, the price of a 20-pack tablet is less as compared to price of 2 packs of 10 pack tablet.
20
In table 2, we observe some price variation in price controlled molecules as well. The price spread in price
controlled molecules come from two sources: a part of the variation is from time series variation, as price
controlled molecules are allowed to revise their prices by a fixed percentage over time to take into account the
rising cost of production. Other part of the price variation is is due to the grouping of SKUs as, per dosage
injections are priced higher compared to per dosage tablets. Note that even though hydroxychloroquine is
under price control, the average price is higher compared to other price controlled drugs. This is because,
hydroxychloroquine is mostly available in injection form, and controlled price for injections is relatively
higher compared to tablets.

10
ufacturing plant till the warehouses is borne by the manufacturers. From the warehouses,
the drugs are delivered to retailers through stockists. Stockists serve as the link between
manufacturer and retailer as well as other institutions (like hospitals) where manufacturer
supplies drugs. Depending on the expected demand, a firm hires stockists to fulfill its need.
For a given region, a firm operates through regional managers as well as several area sales
managers. Area sales managers maintain the drug distribution network through indepen-
dent sales representatives as well as medical representatives to educate the doctors about
the drugs. Hence, in our analysis, the costs involved for the manufacturer include costs of
medicine transportation to the regional warehouses, handling & delivery, obsolescence costs,
capital costs, costs of promotion & education as well as costs of hiring the stockists and
retailers. Storage costs also play an important role in India and vary significantly across
regions. For example, injections are required to be stored in cold storages, which requires
uninterrupted supply of electricity. Since availability of electricity varies significantly across
regions fixed costs also vary across different regions.

3 Model
To evaluate welfare implications of different price control policies, we need to model how
firms endogenously adjust the set of products in response to changing market conditions.
This section presents a two stage model that describes firm’s decisions on offering a set of
products in the market, setting prices of the offered products as well as consumers’ decisions
on choosing those products.

3.1 Demand
We follow BLP and model demand for drugs by a random-coefficient-logit specification.
A market is defined as a region-quarter combination. We denote a market by rt where r
stands for region and t for time period. In a market, multiple firms sell malarial drugs, and
we denote the combined set of available products in the market by Jrt . A malarial patient
chooses at most one of these products, or chooses the outside option of not purchasing any of
them. In our empirical application, the outside option includes treatments using traditional
methods as well as treatments available in public hospitals.
We assume that a consumer chooses one of the offered drugs in the market to maximize
her indirect utility function. Since number of tablets or injections prescribed for a complete
dosage for treatment of malaria varies by molecules, the choice of a consumer in our context
essentially refers to the prescribed amount of the drug for completing a course while treating

11
a malarial patient (for example, 500 mg of Chloroquine in total over 3 days).21 Following
discrete choice literature, the utility specification derived by consumer i from consuming
drug j sold by firm f in market rt is given by
uif jrt (xf jrt , pf jrt , ξf jrt , νirt , Zirt ; θd )
=xf jrt β − α1 pf jrt 1 {r ∈ Rich Region} − α2 pf jrt 1 {r ∈ Poor Region} + ξf jrt
(3.1)
K
X
+ [σ k νirt
k
+ λk Zirt
k
]xkf jrt + εif jrt
k=1

In the above specification, price of drug j by firm f in market rt is denoted by pf jrt . To

capture the differences in taste for price across regions in India, we allow price coefficient
to vary depending on rich and poor regions,22 and denote the coefficients by α1 and α2
respectively. xf jrt is a K-vector characteristics of the drug j by firm f in market rt observed
by the econometrician. In our specification, in xf jrt we include number of SKUs offered by
the firm in market rt in solid form and in liquid form for the given molecule, number of
years since a brand is active in a given region, as well as a measure of presence of firm in
other close therapeutic categories.23 To capture consumer’s average taste for a brand, we
include brand (firm-molecule specific) fixed effects. To capture general time trend across
regions, we also include region-year fixed effects. Additionally, to capture seasonality, we
include quarter fixed effects. We expect number of SKUs to affect utility positively, as
consumers value different types of presentations. For example, children can intake malarial
drugs mostly through liquid form, and hence having more variety may increase utility from
a product. The effect of age of a brand in a given region on utility can be ambiguous. On
one hand, it is possible that patients value old drugs less, on the other hand, given brand
loyalty and trust, drugs may command higher market share with time. We expect firm’s
presence in related therapeutic categories like cold, fever and antibiotics to affect the utility
for malarial drug positively. If a firm is popular in those related categories, then the firm
may also enjoy more brand-recognition and trust among the consumers in the malarial drug
market. The variable ξf jrt is a demand shifter not observed by the econometrician, but
k
observed by the consumers while purchasing the products. νirt is K-vector standard normal
variable assumed to be IID across consumers as well as across product characteristics and
21
For dosage information, we referred pharmacological literature; the details and references are provided
in appendix A.1.
22
We divide all regions into rich and poor regions based on average per capita income. We treat median
of per-capita income across all regions as the cut-off value. Regions, where average per-capita income that
are above this cut-off are treated as rich regions and rest are treated as poor regions.
23
The measure of presence is given by revenue share of a firm in fever, common cold and antibiotics drugs
segments. Note that, these are drugs used to cure parasitic diseases, and are treated as close therapeutic
categories to malaria by the firms.

12
 k
price. Similarly, Zirt denotes the demographic variables drawn from empirical demographic
distribution data (consumption survey data).24 To capture heterogeneity of preferences
across consumers, we allow five random coefficients for interaction of product characteristics
with consumer income. We allow for heterogeneity of consumer preferences for age of a
brand in a region, as well as the number of varieties of products offered by the firm (number
of SKUs). We include interaction of dummy for old molecule with consumer income to
capture heterogeneity in preferences for old molecules. Similarly, allowing interactions of
new molecule dummy in rich regions and in poor regions with consumer income, we capture
the heterogeneous responses of consumers for new molecules with change in income. Finally,
εif jrt denote taste shifters that are assumed to be IID across consumers and across products.
Following literature, εif jrt are assumed to follow Type-I Extreme value distribution. We
denote the demand parameters as θd = (β 0 , α, σ 0 , λ0 ). This specification yields the familiar
logit choice probabilities for a product for each consumer. After integrating out over the
total number of simulated consumers, ns, we arrive at the market share for product j.

3.2 Supply
We model supply decisions of the firms as a static two stage game of complete information.
In a given period, each firm has a menu of potential products to sell across different regions
in India. In the first stage, firms simultaneously choose the set of products to offer across all
regions, and incur those fixed cost payments. Firms commit to the set of released products,
and in the second stage simultaneously choose prices for all the offered products.
In our application, we need to decide on the set of potential firms in a region as well as
the menu of potential products for a given firm. We observe a stable set of firms making
drug offering and withdrawal decisions across different regions in India. In our sample, firms
vary in their sizes and regions of operation; some firms have national presence and others
are regional firms with operations across a set of regions. We include any firm that sells a
product (not just malaria related product) in a region as a potential firm for that region.
A firm specific ‘menu of potential products’ includes any malarial related product that the
firm offers during our sample. Molecules for treatment of malaria are typically developed by
going through a complex R & D process. Also, as already discussed, R & D for molecules
sold by firms in Indian market are funded by external sources and all the firms in India sell
generic drugs for malaria. Furthermore, during our sample period, we do not observe arrival
of any new molecules. Hence, we treat the menu of potential products for each firm to be
exogenously given and fixed for the entire sample period. Next we describe the details of
24
The consumption survey data is obtained from National Sample Survey Organization in India.

13
the game.

3.2.1 Pricing: Stage 2

The second stage decision of the firm involves setting the prices for the products that were
released in stage 1. In the beginning of stage 2, each firm f observes the realization of
demand and cost shocks (ξf jrt , ωf jrt ) for each product j chosen in stage 1. These shocks are
not observed by the econometrician. With the knowledge of these shocks, firms set prices
for each product j, simultaneously in a complete information framework, with the goal to
maximize profits.25 In our model, even for a product not under price control, each firm sets
the same price in a given time period for a given product across all regions in India where
the firm chooses to offer the product.
Let us denote the set of regions where a firm f sells any product at time t by Rf t . For
any region r ∈ Rf t , let us denote the set of products offered by firm f as Jf rt . The profit
maximization problem of a multi-product firm is then given by
  
 X X 
max πf t =  (pf jt − mcf jt )sf jrt (pt ) × Mrt − Total Fixed Costf rt  (3.2)
pf jt ,j∈Jf t  
r∈Rf t j∈Jf rt

where pf jt is the price charged by the firm f for drug j at time t across all regions where the
drug is offered by firm f . Mrt denotes the market size for region r at time t. sf jrt denotes
the equilibrium share of drug j by firm f in region r and time t. mcf jt is the marginal cost
associated with product j by firm f at time t. Note that sf jrt depends on pf jt , as well as on
the entire vector of prices of the products offered across different regions in India at a given
time t. We assume that given any stage 1 history and any parameter values, stage 2 prices
are determined26 uniquely in a pure strategy, interior Nash-Bertrand price equilibrium.
We model the log of marginal cost for a drug j by firm f in a time period t to depend
linearly on the observed cost shifters, wf jt and on an additive error term ωf jrt :

log(mcf jt ) = wf jt γ + ωf jrt (3.3)

where γ is the parameter vector to be estimated. Since these products are generic drugs
with well-known technologies, we assume marginal cost to remain unchanged with level of
production. We include molecule dummies for all 11 molecules to capture cost differences
across molecules. To allow for differences in cost of production of big firms from small firms,
25
Note that, in our application, a product is defined as firm-molecule-sku type (solid/ liquid).
26
Following literature (Fan (2013), Eizenberg (2014)), we assume that a pure strategy Nash equilibrium
exists. Finding a set of sufficient conditions for the existence of Nash equilibrium is beyond the scope of this
article.

14
we include a dummy for big three firms. Since branded products also maintain higher quality,
we expect higher cost of production for big three firms. We also include number of solid and
liquid varieties (SKUs) offered in the market as explanatory variables and expect cost to be
higher for a product with higher number of varieties. Finally, we include presence of firm in
close therapeutic categories and number of years since the brand is active, as observable cost
shifters (wf jt ). We expect the presence of firm in a given region to have negative correlation
with cost of production, as a firm with presence in multiple categories may enjoy better
returns to scale.

3.2.2 Product Offerings: Stage 1

In the first stage, firms observe the realization of fixed cost shocks and make product offering
decisions with the understanding that their actions and their rival’s actions will affect the
variable profit in the second stage. This leads to strategic interaction among firms while
making product offering decisions.
Each firm is assumed to have a pre-specified menu of products and each product has
associated fixed cost which the firm would incur conditional on offering the product in the
market. Our fixed costs include distribution costs, electricity, insurance and other storage
costs, obsolescence costs, costs of promotion & education as well as costs of hiring the
stockists and retailers. Additionally, for a multi-product firm, the fixed costs also include
the opportunity costs of offering the product. These costs are expected to vary across
regions and firms. For example, some firms may have better arrangements in a given region
for electricity compared to other firms. Similarly, storage and insurance costs for injections
can be very different from tablets, as injections may need more care and better refrigeration
facilities with uninterrupted electricity supply. Additionally, in the drug supply chain, each
firm operates in a given state by hiring stockists for supplying the products to the retailers.
These stockists while signing contracts charge differentially for each product, as their charges
typically vary with the form and brand of the drug. Hence, in our specification, we allow for
a flexible fixed cost structure and let fixed cost to vary by firm, product and region.
For each product j offered by firm f in market rt, the fixed cost is assumed to take the
following specification

Ff jrt = Wf jrt θ + νf jrt

(3.4)
{1(region = r)} θ0r + θ1r Wf1jrt + θ2r Wf2jrt + νf jrt
X

=
r

where θ is the vector of fixed cost parameters to be estimated. Wf jrt are fixed cost covariates
which include region specific dummies, as well as region-firm-product specific covariates

15
that includes interaction of firm presence and number of SKUs offered by the firm for a
molecule (Wf1jrt ), and age of the brand in a region (Wf2jrt ). The part of the fixed cost that is
unobservable to the econometrician is denoted by νf jrt . Note that νf jrt is observable to the
firm while making the product offering decision.
Given our information structure, firms have complete information regarding product
specific fixed costs for each potential product. However, while making product offering
decisions, firms are assumed to know the distribution of demand and cost shocks (Fξ , Fω ).
They observe the realization of these shocks only after stage 2 is realized, after having
committed to the set of products to be released. In our demand estimation, we control for
systematic brand effects as well as region-time effects using various fixed effects. Hence, our
assumption here implies that the demand and supply shocks do not capture any systematic
effects that the firms are likely to know prior to committing to their product choices. Such
timing assumptions are also made in, for example, Eizenberg (2014), Wollmann (2017), and
Fan and Yang (2016).
Given our setting, while taking product offering decision, firms form an expectation over
the shock distributions to compute the hypothetical expected profits from any set of product
offerings. We denote by Jf,rt the set of products offered by firm f in market rt. We denote
the variable profit by the firm from offering the product portfolio {Jf,st }s∈Rf t by


π {Jf,st }s∈Rf t , {J−f,st }s∈Rf t ,

where {Jf,st }s∈Rf t denotes the set of products offered by firm f across all regions (Rf t ) at
time t, and {J−f,st }s∈Rf t denotes the set of products offered by the rivals (all firms except f )
across relevant regions at time t. The expected variable profit is given by


Eξ,ω πf rt {Jf,st }s∈Rf t , {J−f,st }s∈Rf t , x, w, p; β, γ, Fξ , Fω =
(3.5)
Z


πf rt {Jf,st }s∈Rf t , {J−f,st }s∈Rf t , x, w, p; β, γ, ξ, ω dFξ dFω
ξ,ω

Firms weigh the expected variable profit from different product combinations against the
total fixed cost of selling the set of products and offer that set of products that maximizes
total expected profit of the firm. Once the product offering decisions are made, firms commit
to these decisions and incur fixed costs for these products.
A Subgame Perfect Nash Equilibrium consists of product choices and prices which con-
stitute a Nash equilibrium in every subgame. However, given that this is a game of complete
information, we might end up with multiple equilibria.27 We follow PPHI, and use necessary
27
In several empirical applications in entry literature, multiplicity of equilibria is handled by assigning a
equilibrium selection mechanism and selecting an equilibrium and assuming that the data is generated under
the assumed equilibrium (For example, see Bajari, Hong, and Ryan (2010)). In our application, we do not
assume uniqueness of equilibrium nor we select any equilibrium from multiple possible equilibria.

16
conditions for equilibrium product selection to estimate entry parameters. These conditions
lead to partial identification of entry parameters and we use tools from moment inequality
literature to do the estimation and inference.

4 Estimation
4.1 Estimation of Demand Parameters
Our demand estimation procedure is similar to that in Nevo (2000a,b, 2001). We estimate de-
mand parameters by minimizing generalized method of moments objective function based on
the conditional independence of demand shocks with respect to the product characteristics,
that is E(ξ|X, Z) = 0. This mean independence assumption is very similar the assumption
made by BLP, except that in our case, independence is assumed for characteristics of all
potential products, rather than the subset of population of products actually offered to the
consumers. Following the literature, our instrumental variables are based on characteristics
of products of the same firm and products of competing firms. Since firms fix same price for
a product across all regions in India, for a given firm we use product characteristics of the
same firm as well as rivals across all regions in India at a given time period while constructing
our instruments.28 The validity of our estimation strategy relies on the timing assumption
that firms do not know demand shocks ξjrt , while they choose product characteristics. Such
timing assumptions are made in, for example, Eizenberg (2014), Wollmann (2017), and Fan
and Yang (2016). In our demand estimation, we control for systematic brand effects as well
as region-time effects using various fixed effects. Hence, though imperfect, it seems reason-
able to assume that any product-time specific shocks are uncorrelated with contemporaneous
product characteristics. Note that our estimation of demand does not rely on supply side
moments. The heterogeneity in demographic characteristics across regions in India and vari-
ation in choice sets due to entry and exit of products over time allows us to take care of
price endogeneity as well as to identify the random coefficients in demand estimation using
BLP instruments.
28
One well-known potential problem while using these instruments is that many of these instruments are
highly correlated both within and across products. Following the suggestion in Conlon (2013), we reduce our
instruments into 14 principal components. These principal components span at least 99% of the variance of
the original instruments, and have the added benefit of forming an orthogonal basis.

17
4.2 Estimation of marginal cost parameters
Much of the prior work has used first order conditions under Nash-Bertrand price equilibrium
to recover marginal costs from prices and elasticity estimates using the demand model. Our
estimation here has two additional complications. First, out of eleven molecules in our
analysis, three29 are under price control. Second, firms set same price for a given product
across all regions in India at a given point in time.
To address price controlled molecules, we recognize that under price control regime, prices
were fixed by using the rule [p = mc × (1 + marg%)]. Hence we recover the marginal cost for
these molecules directly from the prices by setting marg = 8%.30 To estimate marginal costs
for drugs not under price control, we use the first order conditions of the profit maximization
problem under an interior Nash-Bertrand price equilibrium. In our setting, a firm charges
identical price for a given product across all regions where the firm operates (Rf t ) at a
given point in time. Hence, the profit maximization problem of the firm considers setting
prices that maximizes the joint profit across all regions and is given by equation 3.2. We
c nc
partition Jf,rt , the set of products offered by firm f in region r at time t into Jf,rt and Jf,rt
denoting products under price control and products not under price control respectively.31
For a product j not under price control, firm f chooses price pf jt to maximize profit and the
corresponding first order condition is given by

∂πf t X ∂sf jrt X

= (pf jt − mcf jt ) Mrt + sf jrt Mrt
∂pf jt r∈R ∂pf jt r∈R
ft ft
  (4.1)
X X ∂sf krt X X ∂sf lrt pf lt ∗ marg%
+ Mrt (pf kt − mcf kt ) + Mrt =0
r∈R k∈J nc
∂p f jt
r∈R l∈J c
∂p f jt 1 + marg%
ft f,rt ft f,rt

The first two terms in the above expression are own price partial derivatives that capture
the response of profit from product j sold across different regions wrt price of product j.
Third term captures the marginal effect of pf jt on profits from other products offered by firm
f across different regions at time t that are not under price control. Fourth term captures
the marginal effect of pf jt on profits from the products that firm f offers at time t that are
under price control.
Rearranging the first order conditions, we can present the set of equations as specified in
4.1 in vector form, given by
29
Price controlled molecules: Chloroquine; Pyrimethamine-Sulphadoxine; Hydroxychloroquine.
30
Similar approach is also used in Dutta (2011) while dealing with price controlled molecules.
31
Note that, given prices are same across all regions, the estimated marginal costs are also same across
regions. Effectively, we include all region specific costs like transportation costs, storage costs, cost of hiring
local medical representatives, electricity costs, and insurance costs in our fixed cost.

18
 ~ = (~p − markup(θd , Xf jrt , νirt , Zirt ))
mc (4.2)

Since [markup(θd , Xf jrt , νirt , Zirt )] is identified given the demand estimates, mc
~ is estimated
as a residual from the price vector. The above expression is a modified version of the familiar
first order conditions that much of the prior work has used for estimation of marginal costs.
The details of the estimation are discussed in appendix A.2. After recovering mc, we follow
the regression equation 3.3, and estimate the marginal cost parameters (γ).

4.3 Estimation of fixed cost parameters

Identification and estimation of fixed cost parameters relies on the variation in the set of
products offered by different firms in a region in India with changing demand. Intuitively, if
we observe firms offering, withdrawing and re-offering malarial drugs with change in variable
profit (due to change in demand), only then we can infer that firms are solving a problem
similar to the one presented in section 3. We can then recover fixed costs as the average
threshold value of a firm’s variable profit while making a product available in a market. To
gain some evidence whether firms adjust their product offerings with changing demand con-
ditions, we refer to figure 1. In figure 1 on the x-axis, we plot the deviation (standardized)32
of rainfall in a given region in a specific quarter. Since average rainfall varies across regions,
we make the standardization to keep the numbers comparable across regions. On the y-axis,
we plot deviation (standardized)33 of number of products offered in a given region in a specific
quarter. Higher rainfall leads to more mosquitoes and higher incidence of malarial cases. As
figure 1 suggests, with increase in demand for malarial drugs associated with higher rainfall,
we observe that the number of products offered in the market by the firms also goes up.
Since we observe endogenous product choice decisions (offering, withdrawing and re-offering
of products) by a (stable) set of firms in a region over time, we use this variation (along
with the structure of a complete information game) to identify and estimate the fixed cost
parameters. Next, we explain the details of the moment inequalities and estimation of fixed
cost parameters.
− Avg. Rainfall in the region
32
Standardized Rainfall Deviation = Rainfall in region-quarter
Avg. Rainfall in the region
33
Standardized Deviation in number of products=
No. of products in region-quarter − Avg. Number of Products in the region
Avg. No. of Products in the region

19
 Figure 1: Plot of product entry and exit with rainfall

Variation of Number of Products wrt Rainfall across Regions

0.4

0.3
Deviation from Region-wise Average producrs offered

0.2

0.1

0
-1.5 -1 -0.5 0 0.5 1 1.5 2

-0.1

-0.2

-0.3

-0.4
Deviation from Region-wise Average Rainfall

Notes: In this figure, in the x-axis, we plot rainfall deviation given by (rainfall in a region-quarter - mean
rainfall in the region)/ (mean rainfall in the region). In the y-axis, we plot deviation in the number of products
given by, (number of products in the region-quarter - mean number of products in the region)/(mean number
of products in the region)

4.3.1 Inequalities, Endogeneity problem, and Solution

The estimation strategy relies on the necessary equilibrium condition that, any unilateral
deviation by a firm from the set of offered products is unprofitable in expectation. We
denote the information set of firm f at the beginning of stage 1 as If t . Firm f ’s information
set contains all relevant demand side ({Xf jrt , Zirt , sf jrt }∀f,j,r ; θd ; Fξ ), as well as supply side
({wf jrt }∀f,j,r ; γ; Fω ; {Wf jrt }∀f,j,r ) information that allows the firm to compute the expected
profit at time t. We denote by Jf,rt the set of products offered by firm f in market rt. Firm
f may operate across a set of regions at time t, which we donote by Rf t . We denote by Jfp ,
the potential set of products by firm f which is assumed to be fixed and exogenously given.
We denote the variable profit by the firm from offering the product portfolio {Jf,st }s∈Rf t by


π {Jf,st }s∈Rf t , {J−f,st }s∈Rf t

20
where {Jf,st }s∈Rf t denotes the set of products offered by firm f across Rf t regions at time t,
and {J−f,st }s∈Rf t denotes the set of products offered by the rivals (all firms except f ) across
relevant regions at time t. Denote indicator dj = 1 when product j is offered in the market
and dj = 0 when j is not offered in the market.
Consider a product j offered by firm f in market rt, i.e. j ∈ Jf,rt . The necessary
condition for Nash equilibrium implies that a deviation by firm f that eliminates j must not
be profitable, that is, firm f ’s saving in fixed costs from not offering j can not exceed the
expected drop in corresponding variable profit. This implies when dj = 1, profit inequality
is given by (Ff jrt ≤ ∆πu ), where ∆πu is given by
h
 i
Eξ,ω π {Jf,st }s∈Rf t , {J−f,st }s∈Rf t − π Jf,rt \ j, {Jf,st }s∈Rf t \r , {J−f,st }s∈Rf t If t


(4.3)

In the expression, Eξ,ω (.) denotes the firm’s expectation over the true distribution of demand
and supply shocks associated with all products.34 Similarly, for a product j present in firm
f ’s menu of products, but not offered by firm f in market rt, j ∈ Jfp \ Jf rt , it must be


that given dj = 0, profit inequality is denoted by [Ff jrt ≥ ∆πl ], with ∆πl given by
h
 i
Eξ,ω π Jf,rt ∪ j, {Jf,st }s∈Rf t \r , {J−f,st }s∈Rf t − π {Jf,st }s∈Rf t , {J−f,st }s∈Rf t If t . (4.4)

If j was not offered, a deviation that adds j to firm f ’s portfolio must not be profitable,
implying that the added fixed cost must exceed the expected gain in variable profits. The
expected variable profits in the right hand side of the inequalities in 4.3 and 4.4 are com-
puted from estimated demand and marginal cost parameters. The details of computation is
provided in appendix A.4.1. Our fixed cost specification is given by

Ff jrt = Wf jrt θ + νf jrt

Taking the conditional expectation of the bounds in equations 4.3 and 4.4, we have

E(Wf jrt θ − ∆πu dj = 1, If t ) + E(νf jrt dj = 1, If t ) ≤ 0 (For Upper Bound)

 
(4.5)
E(Wf jrt θ − ∆πl dj = 0, If t ) + E(νf jrt dj = 0, If t ) ≥ 0 (For Lower Bound)
 

Further, we assume that expectation of the error term (unconditional on product offering)
is 0, i.e.
E(νf jrt |If t ) = 0. (4.6)
However, if we can also assert that

(E(νf jrt |dj = 1, If t ) = 0) and (E(νf jrt |dj = 0, If t ) = 0) , (4.7)

34
Due to timing assumptions, the firm does not observe the realizations ξ, ω, but has knowledge of the
distributions (Fξ , Fω ) while making the product offering decision.

21
then we can use the (conditional) moment inequalities in 4.5 to identify and estimate upper
and lower bounds on fixed cost parameters. But, note that, given our timing assumption,
the firms observe νf jrt while choosing the offered product portfolio. Hence, the conditional
expectation of νf jrt given the firm choices needs not be zero. In particular, a firm must
have chosen products with favorable fixed cost shocks to offer in the market, and products
with unfavorable fixed cost shocks not to offer in the market. This implies that for a profit
maximizing firm, due to the selection of shocks,
E(νf jrt dj = 1, If t ) ≤ 0, (for Upper bound)

(4.8)
E(νf jrt dj = 0, If t ) ≥ 0, (for Lower bound)


Note that, inequalities for upper bound are derived only for products that are offered in the
market, and inequalities for lower bound are derived only for products that are not offered in
the market. Given our fixed cost specification in 3.4, more negative values of νf jrt imply more
favorable fixed cost. Intuitively, conditions in 4.8 suggests that, for the products selected by
the firms to be offered in the market, the fixed cost shocks must be favorable, which implies
that overall conditional expectation can be negative.35 This selection problem creates an
obstacle in estimation of fixed costs.
To see this, note that even when the upper bound expression in equation 4.5 holds, the
condition
E(Wf jrt θ − ∆πu dj = 1, If t ) ≤ 0


does not necessarily hold as (E(νf jrt |dj = 1, If t ) ≤ 0). If we are willing to ignore this, assume
the conditional expectations to be equal to 0 (as in equation 4.7), and use the corresponding
moment conditions for estimation of parameters, estimates will be biased. Additionally,
during estimation, we may also end up with empty intervals for fixed costs.36 This “selection
problem” is highlighted in PPHI, and Pakes (2010).37 PPHI and Ho and Pakes (2014)
propose an innovative strategy based on “matching pairs” to address this problem. They
match individuals with similar observable characteristics but making different choices, add
up corresponding inequalities for matched agents to derive moment conditions that they can
use for estimation. In our context, to apply this strategy we need to assume that, conditional
on a given observable characteristic (say v), the realized fixed cost shocks from offering and
not offering a product are identical, that is,

(νf jrt |dj = 1, If t , v = ṽ) = (νf jrt |dj = 0, If t , v = ṽ). (4.9)

35
The intuition for lower bound follows similarly.
36
We explain how this can happen using a simple example in the appendix A.3.
37
This problem is also discussed in Ho and Pakes (2014), Eizenberg (2014) and Dickstein and Morales
(2013) among others.

22
Formally, this strategy will control for unobservable heterogeneity in fixed costs only if for
a matched pair, given the observable characteristics, the shocks from offering a product and
not offering a product are identical both in magnitude and sign. However, in our application
this condition imposes strong restrictions. In our case, selection problem arises precisely
because conditional expectations of the shocks take opposite signs (as in equations 4.8).38 39
Our approach borrows ideas developed in Manski and Pepper (2000, 2009) and uses
monotone instrumental variables to solve this problem. For brevity, we first discuss the
solution for the upper bound case. For upper bound, similar to the MIV assumption in
Manski and Pepper (2000) our assumption on νf jrt is yields a mean-monotonicity condition.
MIV assumption: Let V be an ordered set with strictly positive support. Covariate v is
a monotone instrumental variable in the sense of mean-monotonicity, if for all (v1 , v2 ) ∈
(V × V ), with 0 < v1 ≤ v2 ,
E(νf jrt dj = 1, If t , v = v1 ) ≤ E(νf0 jrt dj = 1, If t , v = v2 )
     
(4.10)

where |.| denotes the absolute value function.

In the above expression, a higher absolute value of expected error (which refers to more
negative unobservable fixed cost) implies a favorable fixed cost shock. In contrast with
4.9, we assume that the absolute value of expectation of error terms follow monotonicity
with respect to v. Our assumption states that, as the instrument takes higher values, the
expectation of fixed cost shocks become more favorable. We construct two instruments
for our application. Our first instrument is based on a firm’s presence in closely related
therapeutic categories in a region. In a given region, firms not only sell malarial drugs,
but also sell antibiotics as well as the drugs that cure cold, and fever. Note that in our
application, the set of firms are stable throughout the sample period, we observe entry and
exit of malarial products varying with seasonality. If a firm enjoys higher revenue share in
those closely related drug segments (excluding malarial drugs) compared to its rival, then it
is reasonable to assume that the firm also enjoys better facilities (like better arrangement of
electricity for storage of drugs, better relationship with stockists, retailers as well as doctors,
better network of medical representatives) compared to its rivals. Given this we assume
that, if a firm enjoys a high revenue share in selling these closely related products compared
38
In another novel approach proposed by Eizenberg (2014), identified set is constructed by adding inequal-
ities for missing bounds. For a product that is offered in the market, missing lower bound is replaced by
non-negativity conditions for fixed costs. Similarly, for a product not offered in the market, the missing upper
bound is replaced by a conservative upper bound. Due to heterogeneity of firms across regions, allowing a
conservative upper bound in our application requires us to make strong assumptions.
39
Another creative strategy developed by Dickstein and Morales (2013) deals with this problem by param-
eterizing the distribution of the fixed cost errors.

23
to its rival, then the firm also enjoys more favorable fixed cost shock in expectation. Our
second instrument relies on the assumption that if a firm sells more varieties of products in
one province, then the firm is likely to face favorable cost shocks compared to other firms.
Hence the number of products, or the number of SKUs for a molecule in a region, could
serve as a second monotone instrument.
In addition to the MIV assumption, we make another normalization assumption to
achieve identification.
Normalization Assumption: For a product under price control, the expected total profit,
conditional on being offered in the market is equal to 0. Hence if j is under price control,

[E(∆πjpc − Wfpcjrt θ − νf jrt |dj = 1, If t ) = 0] (4.11)

This assumption is the ‘zero expected total profit’ condition for products under price control
that are offered in the market. Price controlled molecules40 are being sold in the Indian
market since early 1970’s. In contrast to other molecules, these drugs are mostly homoge-
neous generic products with little brand differentiation. Almost every firm in the market is
a potential entrant for those molecules. Additionally, due to price control regime, all firms
selling these drugs charge same price. Hence, under free entry condition and homogeneous
products, we assume that the expected total profit from offering these products is zero.41
Using 4.10 and 4.11, we derive the inequalities for estimation of upper bound of fixed
cost parameters. Note that, given 4.10, with 0 < v1 ≤ v2 , we have
E(νf jrt |dj = 1, If t , v = v1 ) ≤ E(νf0 jrt |dj = 1, If t , v = v2 )
   

=⇒ − E(νf jrt |dj = 1, If t , v = v1 ) ≤ −E(νf0 jrt |dj = 1, If t , v = v2 ) (4.12)

(using 4.8 since, E(νf jrt |dj = 1, If t , v1 , v2 ) ≤ 0)

=⇒ − E(νf jrt v1 |dj = 1, If t , v = v1 ) ≤ −E(νf0 jrt v2 |dj = 1, If t , v = v2 )

(since 0 < v1 ≤ v2, multiplication preserves inequality) (4.13)
=⇒ −E(νf jrt v1 |dj = 1, If t , v = v1 ) + E(νf0 jrt v2 |dj = 1, If t , v = v2 ) ≤ 0

Using the expression for upper bound in 4.5, and equation 4.11, we have,
E(∆πu − Wf jrt θIf t , v = v1 ) − E(νf jrt dj = 1, If t , v = v1 ) ≥ 0, and
 

E(∆π pc − W pc θIf t , v = v1 ) − E(νf0 jrt dj = 1, If t , v = v2 ) = 0.

 
j f jrt

40
Price controlled molecules are 1. Chloroquine, 2. Hydroxychloroquine, 3. Pyrimethamine & Sulphadox-
ine.
41
Note that, homogeneous products and free entry does not necessarily guarantee zero economic profit
in the presence of fixed costs. We assume that the expected total profit is zero; however, our identification
remains valid even when expected total profit is positive, but not large enough to exceed the expected total
profit from products not under price control.

24
Multiplying with v1 and v2 respectively and taking the difference,
− E(∆πu v1 − Wf jrt θv1 If t , v1 ) − E(νf jrt v1 dj = 1, If t , v = v1 ) +
 

E(∆πjpc v2 − Wfpcjrt θv2 If t , v2 ) − E(νf0 jrt v2 |dj = 1, If t , v = v2 ) ≤ 0



(4.14)
=⇒ E (Wf jrt v1 − Wfpcjrt v2 )θ − (∆πu v1 − ∆πjpc v2 )If t , v1 , v2
  

≤ −E(νf jrt v1 dj = 1, If t , v = v1 ) + E(νf0 jrt v2 |dj = 1, If t , v = v2 ) ≤ 0.



Given 4.13, the conditional moment inequalities for upper bound are given by

mu (Wf jrt , θ; If t ) ≡ E (Wf jrt v1 − Wfpcjrt v2 )θ − (∆πu v1 − ∆πjpc v2 )If t , v1 , v2 ≤ 0

  
(4.15)

With similar calculations, we can also derive the conditional moment inequalities for lower
bound:

ml (Wf jrt , θ; If t ) ≡ E (∆πl v1 + ∆πjpc v2 ) − (Wf jrt v1 + Wfpcjrt v2 )θIf t , v1 , v2 ≤ 0

  
(4.16)

The derivation of moment inequality for lower bound and details of construction of uncon-
ditional moment conditions are discussed in the appendix A.4.2 and A.4.3. Corresponding
sample-moment conditions are given by:
1 X
m̄k = mk (Wf jrt , θ, If t ); k = 1, . . . , K
n f,j,r,t

where n denotes the sample size, and K denotes the number of sample moment conditions.
Since our fixed cost parameters are region specific, and for each region, there are 3 parameters
to be estimated, we end up estimating 69 parameters for 23 regions in our sample.
To report confidence intervals, we construct sets in which the fixed cost parameters will
uniformly lie 95% of the time. Inference based on inequalities is less straightforward than
inference based on equalities (for example generalized method of moments) since inequalities
provide only one-sided restrictions. We use methodologies developed in Kaido, Molinari, and
Stoye (2016)[KMS] to construct confidence intervals for fixed cost parameters. We report
element wise confidence interval for our parameter vector θ. As KMS points out, the key step
in computation is to get the critical value right. The 95% coverage is achieved by properly
calibrating the critical value, which is computed by checking feasibility of a linear program
across bootstrap repetitions. While calibrating the critical value, we select binding moments
following the generalized moment selection as discussed in Andrews and Soares (2010). The
details of this computation are described in A.4.4.

25
5 Estimation Results
5.1 Demand parameters
Table 3 reports the results from estimation of demand system. The demand parameters
are, with few exceptions, estimated very precisely. Especially, the random coefficients for
interaction terms with individual income are all precisely estimated at acceptable levels of
significance. A few parameters deserve discussion. Price coefficient is negative and significant
both in rich and poor regions. Additionally, in line with expectation, consumers in poor
regions dislike price more compared to rich regions.
[Table 3 about here]
Positive and significant coefficient for number of SKUs suggests that firms can increase
their market share by differentiating themselves in terms of presentations. Given that most
of the consumers pay out of pocket and are poor, having multiple presentations in the market
provides consumers with more options and makes the product more attractive. Similarly,
firm’s presence in close therapeutic categories like cold, fever and antibiotic segments picks
up a positive coefficient. This squares with the hypothesis that familiarity of the brand name
may lead to better brand recognition and reliability among consumers. The preferences for
age of a brand in a region first increases with age and then decreases, suggesting that newly
introduced molecules take some time to gain market share, but consumers prefer newer
molecules over old molecules. Among the random coefficients, the coefficient of income
interacted with the dummy for price control molecules is negative and significant. Similarly,
the dummy for new molecules interacted with income is positive and significant in both rich
and poor regions. Note that, this suggests that as income increases, a consumer derives
more utility from new molecules compared to old ineffective molecules that are under price
control. With increase in income, consumers derive more utility from newer products as well
as from products that are available in more presentations.
[Table 4 about here]
These estimates imply sensible elasticities and markups. Table 4 reports average own
price elasticities for molecules that are not under price control, separately for rich and poor
regions. Demand appears to be elastic with majority of the own price elasticities in poor
region around -4 and in rich region less than -1.5. Interestingly, these numbers are very
similar to the numbers that Chaudhuri, Goldberg, and Jia (2006) report in their study of
antibiotic segment of Indian pharmaceutical market and Dutta (2011) reports in her study
of Indian pharmaceutical industry. Additionally, as expected, demand in poor regions are
more elastic compared to rich regions. These findings are key drivers of our results in

26
the counterfactual exercise. Note that given such high variations in elasticities along with
variation in fixed costs across regions, charging a single price across all regions in India may
lead to significant inefficiencies in drug distribution decisions.

5.2 Marginal cost parameters

Following 3.2, we use the first order conditions from a firm’s joint profit maximization prob-
lem across all regions where it offers the products and compute the implied markup. Using
the equations in 4.2, the marginal costs are estimated as a residual from prices by subtracting
the markups from prices.

[Table 5 about here]

Our markup estimates are reported in table 5. For the three molecules that are under price
control since 1995, we recover marginal cost from prices by imposing 8% markup rule. For
other products markup varies in the range of 20% to 45%. The implied markups in our case
are in the same ballpark of the numbers that Chaudhuri, Goldberg, and Jia (2006)42 recover
while studying the antibiotic segment in India.
After recovering the marginal cost for each product, we regress the log of marginal costs
on a set of regressors. The regressors include the dummy for big three firms, number of
solid and liquid presentations produced by the firm for the molecule, presence of the firm in
related therapeutic categories like cold, fever and antibiotic segments, as well as number of
years since the firm is active in the molecule. Additionally we also include molecule dummies
and quarter dummies in our regression analysis.

[Table 6 about here]

Table 6 reports the marginal cost parameter estimates. All the parameter estimates
are significant at 95% level of significance and R-square is close to 0.9. Marginal cost of
production for the top three firms is higher compared to other firms suggesting quality
variations across branded and unbranded generic drugs. Similarly, cost of producing an
additional presentation of injection and tablet is costly, and production of injections are
more expensive compared to tablets. If a firm is already popular in related therapeutic
categories, it faces less cost of production. The estimates also suggest that the marginal cost
of production is higher for newer molecules and this goes down with time.
42
Their implied markup under monopoly assumption varies from 20% to 60%.

27
5.3 Fixed cost parameters
Table 7 reports results from fixed cost estimation from 23 regions in India. Estimation of
confidence intervals is an expensive computational procedure due to partial identification
of fixed cost parameters. Note that, fixed costs here include distribution costs, electricity,
insurance and other storage costs, obsolescence costs, costs of promotion and education,
costs of hiring the stockists and retailers, as well as the opportunity costs of offering the
product. As size of the states (in terms of area), infrastructural facilities as well as firm-
specific arrangements across the regions vary significantly, fixed costs are also expected to
vary across regions and firms. Hence, to flexibly capture the heterogeneity across states, we
allow fixed costs to vary with region specific covariates; since for each region there are three
parameters to be estimated, for 23 regions we end up estimating bounds for 69 parameters.
In table 7, for each region, we report the confidence-sets of the coefficients of fixed cost
covariates. These bounds are the 95% confidence intervals that uniformly cover the projection
of the identified set and are computed using the projection based inference developed in
Kaido, Molinari, and Stoye (2016). We include a region specific constant in the fixed cost
specification to capture the regional heterogeneity that is common to all firms. In a given
region, interaction of number of SKUs and presence of firm in close therapeutic categories
as a control in fixed cost estimation allows fixed cost to vary with respect to firm size as well
as the number of presentations that a given firm offers. We include ‘number of years since a
product is active in a region’ to allow the fixed cost to vary for a product over time.

[Table 7 about here]

To interpret these estimates, in table 8, we report average fixed cost bounds. While com-
puting the upper and lower bounds for each parameter, we also derive the boundary points
of the confidence interval of the identified set. For a given product, we can then compute
fixed costs (in Indian Rupees) at each of the boundary points, and find the lower and upper
bounds of fixed costs for each product. In table 8, we report the average of minimum and
maximum fixed costs across products in a region. A firm’s product entry and exit decisions
would depend on a comparison of fixed cost with corresponding variable profit. Hence, for
each region, we also provide average fixed cost as a percentage fraction of average variable
profit in table 8.

[Table 8 about here]

These fixed cost estimates imply sensible numbers. Although in table 7, some of the
confidence intervals of parameter estimates include 0, all parameters equal to 0 (i.e. [0, 0, 0]T
) is rejected at 95% level of significance for every region. This suggests that fixed costs for all

28
products equal to zero is rejected for all regions. Note that, fixed costs are allowed to vary
across products in each region and fixed cost bounds are estimated for all potential products
in every market. In close to 95% of the product level fixed cost estimates, bounds are positive
and significantly different from zero; in rest 5% of the cases, we could not statistically reject
the hypothesis that fixed cost equals zero. Our estimates highlight two key findings. First,
fixed costs (as computed in Indian Rupees) are heterogeneous across regions and firms.
Larger regions are expected to have higher fixed costs as more distributional and storage
related spendings are required to supply drugs to different parts of the region. However,
fixed costs also depend on the infrastructural facilities available in the region. For example,
although both Odisha and Gujarat are similar in size (in terms of area of region),43 average
per quarter fixed cost for offering a product in Odisha lies in the interval 209 thousand to 240
thousand Indian rupees, whereas offering a product in Gujarat costs between 24 thousand to
71 thousand Indian rupees. This squares well with the observation that, majority of firms are
headquartered in Gujarat. Additionally, being a more developed state, Gujarat has better
transportation as well as infrastructure facilities compared to Odisha.
Second, fixed costs are significant proportions of variable profits and these ratios44 vary
across regions. As our results in table 8 reveal, in regions like Odisha, Bihar, Chhatisgarh,
Madhyapradesh, where infrastructure facilities are underdeveloped and access to electricity
can be challenging, 45-75% of variable profit is spent on fixed costs. However, in states
like Gujarat and Karnataka, where most of the firms are headquartered and have their
plants located, fixed cost can be relatively small fraction (20-30%) of variable profit. Note
that, our finding of significantly large fixed costs compared to variable profit, is a pivotal
component while studying the effects of price regulation. If on the contrary, fixed costs were
low as compared to variable profits earned by firms, then a profit maximizing firm would
continue offering malarial drugs even when low price caps are mandated by the government.
However, our results suggest that in response to stringent price control policy, firms may
find it profitable to stop offering the products under price control and exit the market.
Three key findings from our estimation (i) heterogeneity of demand characteristics across
regions, (ii) heterogeneity of fixed costs across regions, and more importantly (iii) fixed
costs being significant proportions of variable profits; are crucial elements for evaluating the
price control policy decisions. Note that, once price control is imposed and the price-cap
is identical for all firms across all regions, incentives for the firms to offer products across
different regions will vary significantly. Depending on the size of the market, and fixed costs
of making the drug available, access to drugs across regions will be affected heterogeneously
43
Area of Odisha is 60,119 mile sq, and area of Gujarat is 75,685 mile sq.
44
Ratio of fixed cost with corresponding variable profit.

29
in response to such policies. In addition, the estimated fixed cost bounds are also important
in the discussion of access to newly launched drugs. As Berndt and Cockburn (2014) and
Cockburn, Lanjouw, and Schankerman (2016) point out, policies like price regulation may
result in delay in launch of new molecules in developing countries. Heterogeneity of fixed
cost bounds from our estimates adds to this discussion by suggesting that, even after a drug
is launched in India, since costs of selling the drug are heterogeneous across regions, the
diffusion of benefits of the innovation may vary significantly across regions.

6 Counterfactual Analysis
Using the estimates from our structural model, we now evaluate price control policies to
see how fixing price cap margins at different levels affects product offerings and consumer
welfare. This exercise also enables us to characterize the optimal price control policy. Section
6.1 discusses the background of price control, section 6.2 provides practical details and results
from our counterfactual exercise.

6.1 Background of counterfactual exercise

Understanding the effects of price control policy has significant welfare implications as gov-
ernment of India has been using price control policy extensively with the goal to improve
access to essential drugs. The policy covers all the drugs that are listed in “National List
of Essential Medicines”; number of drugs under government price controls has increased to
652 in 2013, and more than 800 drugs in 2016 including combination products, from just
74 bulk drugs in 1995.45 In our counterfactual exercise, we simulate what would have hap-
pened in terms of product availability decisions and corresponding consumer welfare, had
the government implemented price control on a subset of products during our sample period.
While current discussions on price control in government’s policy-documents focus primarily
on lowering the prices so as to make the drugs affordable, our analysis highlights the role
of firms’ incentives to make the drugs available in local markets in response to price con-
trol, as an important determinant of consumer welfare, which we find missing in the current
policy discussions. Our counterfactual exercise aims at highlighting the fact that, fixed cost
of making a drug available in a market is high enough to induce product withdrawal and
welfare loss to the consumers in response to setting low prices under price control.
45
The estimated size of the total market being affected by this policy is close to USD 13.1 billion which is
approximately 35% of the Indian pharmaceutical market. (Reference - ihs website https://www.ihs.com/
country-industry-forecasting.html?ID=1065979298)

30
 To add some historical perspective, we refer to the fact that, after 1995 price control
policy, bulk drug46 manufactures of 25 out of the 74 price controlled molecules moved out of
India.47 Indian manufacturers rely on import of those bulk drugs for medicine production.
Given this historical fact, a provision in the drug price control act 2013 states the concern of
product withdrawal of the firms from all regions in India, and the policy discourages firms
to discontinue the sale of the product from all states in India.48 While withdrawal from all
parts of India is a concern, our analysis highlights additional welfare loss due to product
withdrawals at the region level.

6.2 Counterfactual Exercise: Details and Results

Motivated by the policy setting, we ask the question: “If price controls were imposed for
a set of products in our sample, how would have product offering decisions and consumer
welfare varied depending on the level of price control?” In our counterfactual exercise, we fix
profit margins at different levels and simulate product offering decisions as well as consumer
welfare.
However, implementing this exercise brings in several computational challenges. Each
potential firm in a given region can produce from a menu of products. Given a menu of k
products, the firm can choose not to offer any product, offer all of the k products, or offer
a subset of products. This gives 2k possibilities for the given firm. Note that, for each of
these combinations, the expected variable profit of a firm will depend on the choices made by
competitors in the market as well, since equilibrium market share depends on choices made
by all firms; for example, share of a product will be less under higher competition and vice
versa. According to our modeling assumption, the firm compares expected variable profit
from each of these combinations of products with the total fixed cost from offering these
products and chooses that combination for which the total expected profit is maximized.
Given that on an average, close to 20 potential firms operate in a region and each firm can
46
Bulk drugs refer to the primary, active ingredient(s) of a final pharmaceutical product, produced in the
first stage of pharmaceutical production and usually in bulk quantities.
47
Source:
https://www.thehindubusinessline.com/opinion/Drug-price-control-does-not-help-consumer/
article20407412.ece
48
To quote price control policy 2013: “Any manufacturer of scheduled formulation, intending to discon-
tinue any scheduled formulation from the market shall issue a public notice and also intimate the government
in Form-IV of schedule-II of this order in this regard at least six month prior to the intended date of discon-
tinuation and the government may, in public interest, direct the manufacturer of the scheduled formulation
to continue with required level of production or import for a period not exceeding one year, from the intended
date of such discontinuation within a period of sixty days of receipt of such intimation.” -[paragraph 21-
DPCO 2013]

31
produce multiple products, it ends up in a very high dimensional problem.49 Additionally,
the policy imposes a price cap for a subset of products, hence for high margin values, a firm
may find it optimal to set a price below the price cap. We need to take this into account in
our exercise. Therefore, to make counterfactual exercise computationally feasible, we make
several choices while implementing the analysis.
First, while simulating product offering, we restrict a firm’s endogenous product choice
decision to three new Artesunate-based molecules. We take the offering decisions of the rest
of the products in a market as given. These three molecules do not display any resistance
to malarial parasite and hence, access to these molecules carry high welfare significance. In
our sample, there are five Artesunate-based molecules. Out of these, we consider Arteether
- Artemotil, Artesunate, and Artemether - Lumefantrine for our analysis, since these three
are the most popular molecules and cover more than 80% of the sales among Artesunate-
based molecules. Second, we only consider the product offering decisions of the top three
firms. These top three firms cover around 70% of the market share of the three considered
molecules. With these choices, for a given region-quarter, the menu of products across three
firms includes 13 choices, leading to 213 possible decisions of the firms.
With these choices, we implement our counterfactual exercise for Odisha region.50 We
choose Odisha for our counterfactual exercise due to its welfare significance, since this state
registers highest incidence of malarial cases and deaths due to malaria in India.51 Addi-
tionally, Odisha is one of the poorest states in India with close to 33% of population below
poverty line as compared to national average of 21%.52 Note that, the goal of our counterfac-
tual exercise is to establish that price control policies leading to low prices may not always
benefit consumers. As consumers in a poor region like Odisha are more price sensitive, we
implement the counterfactual exercise in a region with significant poor population to cred-
ibly establish our claim. We simulate the product offering decisions as well as consumer
welfare at different margin values for two quarters in 2011 and 2012. These two quarters
in 2011 and 2012 register highest number of malarial cases in the region. The margin val-
ues we use for evaluation range from 1%, 2%, 3%, . . ., up to 30%. For each margin value,
for a product under price control, we fix the corresponding price caps by using the rule
[price = mc × (1+margin%) ], where mc denotes the marginal costs estimated from supply
model. For products that belong to molecules other than these three molecules, we take the
49
With 20 firms each producing 5 products on average, the total number of choices are 2100 ∼ 1030 .
50
Restricting our analysis to one region is again motivated by practical computational burden.
51
Source: Dhingra, Jha, Sharma, Cohen, Jotkar, Rodriguez, Bassani, Suraweera, Laxminarayan, Peto,
et al. (2010).
52
Source: Reserve Bank of India report (https://web.archive.org/web/20140407102043/http://www.
rbi.org.in/scripts/PublicationsView.aspx?id=15283).

32
offering decisions as given.
To make things clear, as an example, let us denote the set of potential products from
other eight molecules53 in a market as Yp . Suppose, we observe that out of the Yp products,
Xo products are offered by different firms in the market. While running our counterfactual
exercise, we consider the product availability decisions for these products (Xo offered out
of Yp ) as given. Given those decisions, we simulate the set of Nash equilibrium strategies
for the other 13 products that belong to the three Artesunate-based molecules. To compute
Nash equilibria, out of the 213 possible decisions, our goal is to eliminate the strategies
with profitable deviations. To proceed, for a given level of margin, for a given strategy
set, we first compute equilibrium prices, equilibrium market shares and variable profits for
all the products offered in the market.54 Since price control policy sets caps on the prices,
while computing equilibrium prices we take this into consideration and solve a constrained
optimization problem. This implies, while solving equilibrium prices, for a product under
price control, firms can charge any price less than or equal to price cap, while no such
constraint applies to products not under price control.
For a given market, for a specific margin value, we run this computational exercise for
all the 213 strategies. For the two quarters in 2011 and 2012, we perform these calculations
for all the 30 different margin values. Note that, if we do not allow other molecules to be
present (that is assume |Xo | = 0) while running our counterfactual, products from these
three molecules will face less competition, and our model will generate more variable profit
for these products. Additionally, lack of substitutes will also overestimate consumer welfare
gain from these three molecules. In the ideal exercise, we should allow all products to enter
the market through endogenous product offering decision. However, due to computational
infeasibility we (unsatisfactorily) resort to this approximation.55
After computing the variable profit from different strategies, we compute the set of coun-
terfactual equilibria in each market, and use this to place bounds on consumer welfare.
Given that our model assumes a complete information framework and no equilibrium selec-
tion mechanism is imposed, we end up with multiple Nash Equilibria. Additionally, due to
partial identification of fixed costs, we end up computing the set of outcomes that can not
be ruled out by the fixed cost bounds as equilibria of the game. We call this set of equilibria
53
Note that there are 11 molecules in total, and offering decisions for three molecules are modeled endoge-
nously.
54
For example, for a strategy (1, 0, 0, 0, 1, 1, 0, 0, 0, 1, 0, 0, 0): firms 1 and 3 produce one product each,
| {z } | {z } | {z }
firm 1 firm 2 firm 3
and firm 2 produces 2 products. Taking into account the Xo products from other eight molecules, in total
(|Xo | + 3) products are offered in the market under this strategy, where |.| denotes the cardinality of a set.
55
We have simulated results where we only consider three molecules and do not allow any other molecules to
be present. These results are available upon request. The qualitative results are robust to this specification.

33
as potential equilibria.56 While computing the set of potential equilibria, we eliminate a
strategy by checking if a profitable deviation for that strategy exists. Specifically, given a
strategy, dropping a product offered under the strategy will be a profitable deviation, when
the saved fixed cost from dropping the product exceeds the loss in variable profit, leading
to increase in overall profit. Hence, when we check whether dropping a product may lead
to a profitable deviation, we consider lower bound of the fixed cost. This is because, when
dropping a product leads to higher overall profit with lower bound of fixed cost, any other
value of fixed cost inside the bounds would also lead to higher overall profit. Similarly, to
check whether adding products may lead to profitable deviations, we consider the upper
bound of the fixed cost. Note that, the set of strategies that survive the elimination process
includes all strategies that qualify for Nash Equilibrium and also potentially includes strate-
gies that are not Nash Equilibrium under the knowledge of true fixed costs. For a given
level of margin, we compute the set of potential equilibria in a given market, and compute
corresponding consumer welfare for each of these strategies in the equilibrium set.57
The goal of our counterfactual exercise is to establish that price control policies with low
prices may actually hurt consumers. This may happen, if firms respond to these policies by
not offering products in the market, resulting in reduced access and lower welfare. Figure
2, figure 3, and table 9 report the results from counterfactual exercise in Odisha. In figure
2, the X-axis plots different margin values and the Y-axis reports the number of products
offered in equilibrium at the corresponding margin values. For a given level of price control
margin, we report the average number of products offered in the market at equilibrium
where the average is taken across all potential equilibria that survive after elimination of the
strategies with profitable deviations. Additionally, for each margin value, we also compute
the maximum and minimum number of products offered across all potential equilibria in
the two quarters in 2011 and 2012 and report as confidence bars. Figure 2 shows that the
average, minimum as well as maximum number of products offered increases with higher
price control margin. Given that offering a product in a market incurs fixed cost, at low
margin values (such as 1% and 2% margin) no products are offered. As margin values
increase, number of products offered across equilibria also increases. Additionally, figure 2
56
Eizenberg (2014) also uses the same nomenclature.
57
Consumer surplus of consumer i is computed following Small and Rosen (1981). For consumer i with
indirect utility function vij takes the following form:
  
J Z
1  X
CSi = γ + ln 1 +
 exp(vij ) ;
 CS = CSi dFZ dPν
αi j=1

where γ is Euler’s constant, αi denotes the price coefficient for consumer i, and vij is the deterministic
component of utility for person i from product j.

34
 Figure 2: Number of products offered across different margin values in Odisha

NUMBER OF PRODUCTS OFFERED ACROSS EQUILIBRIA

8

6
Number of Products Offered

0
1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 11% 12% 13% 14% 15% 16% 17% 18% 19% 20% 21% 22% 23% 24% 25% 26% 27% 28% 29% 30%

Price Control Margin Values

Notes: In this figure, we plot margin values from 1% to 30% in the x-axis. The y-axis plots the number of
products offered in equilibrium at different margin values. The confidence intervals show the maximum and
minimum number of products offered in the market across all potential equilibria.

reveals that the number of offered products increases at a faster rate from 1% to 16%. After
16%, the number of products offered across equilibria either remains constant or increases
marginally.

[Table 9 about here]

Table 9 reports minimum and maximum number of products offered for the three molecules
separately across different margin values. For example, the entry in the table for 1% margin
shows that, at price control with 1% margin, no products are offered for any of the three
molecules. The table reveals access to different molecules across price control margins. In
particular, the table shows that, for price control margins up to 9%, none of the firms offer
any drug belonging to the molecule Artemether-Lumefantrine, leading to no access to this
molecule at low prices. Even when the margin is set at 14%, some of the equilibria may lead

35
to no drugs being offered for this molecule. Access to all the molecules is ensured only when
prices are high enough (in this case at least 15% margin).

Figure 3: Consumer welfare (in INR) across different margin values in Odisha

C O N S U M E R W E L FA R E A C R O S S D I F F E R E NT M A R G I N VA L U E S

1,030,000

1,025,000

1,020,000
Consumer Welfare in INR

1,015,000

1,010,000

1,005,000

1,000,000

995,000

990,000

985,000
1% 2% 3% 4% 5% 6% 7% 8% 9% 10% 11% 12% 13% 14% 15% 16% 17% 18% 19% 20% 21% 22% 23% 24% 25% 26% 27% 28% 29% 30%

Price Control Margin Values

Notes: In this figure, we plot margin values from 1% to 30% in the x-axis. The y-axis plots the average
consumer welfare at different margin values at equilibrium. The confidence intervals report the maximum
and minimum consumer welfare across multiple potential equilibria at different margin values.

Although higher margin results in increased access to drugs, leading to increased con-
sumer welfare, higher margin also implies higher price, which puts negative pressure on
consumer welfare. To see which of the two forces dominate and compare across margin
values, we refer to figure 3. The X-axis in the figure plots different price control margins
and the Y-axis reports the corresponding average consumer welfare, where average is taken
across all potential equilibria. Average consumer welfare can be interpreted as the expected
attainable consumer welfare level, when each of the equilibrium inside the set of multiple
potential equilibria has equal probability of getting realized. Additionally, in figure 3, for a
given margin value, the confidence interval bars present the maximum and minimum con-
sumer welfare across all potential equilibria. Note that, for some margin values we end up

36
with unique equilibrium and hence, the maximum, minimum as well as average welfare level
coincide with each other. As figure 3 demonstrates, consumer welfare follows an almost
inverted-U shaped pattern with respect to different margin values. As margin is increased
from 1% to 16%, since number of products offered increases at a fast rate, average consumer
welfare increases, even when price increases. Beyond 16%, negative pressure due to increase
in prices dominates the positive effect due to higher products and the average consumer
welfare goes down. Hence, in terms of average consumer welfare, 16% profit margin qualifies
as the optimal price control rule. This result is striking, as it suggests that, even in a poor
region like Odisha, charging higher price by almost doubling the margin as compared to gov-
ernment’s price control rule (that sets margin close to 8-10%), leads to increased consumer
welfare. Note that due to partial identification of fixed costs and multiplicity of equilibria,

Figure 4: Consumer welfare comparison at 8% and 16% in Odisha

C O N S U M E R W E L FA R E C O M PA R I S O N AT 8 % A N D 1 6 %
1,025,000

1,024,500

1,024,000

1,023,500

1,023,000

1,022,500

1,022,000

1,021,500

1,021,000
Max at 8% Min at 16%

Notes: In this figure, we plot the maximum consumer welfare across all potential equilibria at 8% and
minimum consumer welfare across all potential equilibria at 16%. The figure shows that even the worst
attainable equilibrium at 16% derives higher consumer welfare compared to the best attainable equilibrium
at 8%.

for each margin value, we recover upper and lower bounds on consumer welfare. Hence, in
addition to comparing average consumer welfare at 16% and 8%, we also consider the min-
imum consumer welfare at 16% margin and compare it to maximum attainable consumer
welfare at 8% margin. As figure 4 shows, the welfare at the lower bound at 16% exceeds the
upper bound of consumer welfare at 8%. To summarize, our results suggest that a stringent
price control policy with low prices may lead to lower consumer welfare even in a poor re-
gion like Odisha, as firms would respond by withdrawing products from the market. This

37
suggests that an optimal price control policy that balances product availability with higher
prices would lead to maximum consumer welfare.
It is worth pointing out here that, documented evidence from China, South Korea and
Philippines also point to the fact that in those countries, price control led to lowering rather
than increasing of access to drugs. In this context, our results suggests that designing
an optimal price control policy need to take both consumer preferences as well as firms’
incentives into account. Ignoring the supply side responses may lead to suboptimal policy
design and may result in unintended consequences for consumer welfare.

7 Concluding Remarks
This paper examines the consequences of imposition of price control on access to drugs and
consumer welfare in the context of regional markets for malarial drugs in India. Incentives
for innovation and welfare loss due to resulting higher prices have dominated the discussions
in the debate for access to drugs. This paper contributes by showing that, even when no
molecules are under patent protection, ‘incentives for making drugs available’ can be crucial
in ensuring that the consumers benefit from price control. Our study helps us to establish
these findings and characterize optimal price control level, as we allow firms to endogenously
adjust product choice decisions across markets in response to changing demand conditions.
Compared to existing studies on pharmaceutical markets in developing countries, we
present a detailed model of differentiated product cost and demand along with regional
variation using firm-level micro-data. Our estimates reveal demand characteristics such as
elasticities vary across regions. Additionally, fixed costs are heterogeneous across regions
and firms, and more importantly, fixed costs are significant proportions of variable profits.
Our counterfactual exercise establishes these finding as crucial determinants for evaluating
welfare implications of the price control regulations.
This study has implications for public policies especially in low-income countries, where
price control policies are widely implemented. Regional heterogeneity in terms of infrastruc-
tural facilities as well as demand characteristics are common features observed in low-income
countries. Therefore, while designing government policies in pharmaceutical markets in those
countries, distributional issues need to be addressed. Our results suggest that, charging op-
timal price control level can be one solution to this issues. In addition, other instruments
such as, providing subsidies to firms for distribution of drugs and even public participation
in drug distribution can be potential ways to address this problem.

38
8 Tables

Table 1: Average Price and Quantity Sold (by Molecule)

Molecule Average Price per patient No of Patients

(Indian Rupees) (in Million)
Price Control from 1995
Pyrimethamine, Sulphadoxine 6 27.6
Chloroquine 9 158.6
Hydroxychloroquine 56 14.9

Not under price control

Quinine 201 4.0
A, S, P 210 0.8
Mefloquine 267 0.3
Arteether, Artemotil 154 17.6
Artesunate 198 7.8
Artemether, Lumefantrine 191 5.9
Artesunate, Mefloquine 457 0.4
Arteether, Lumefantrine 275 0.3
*A, S, P: Artesunate, Sulfadoxine, Pyrimethamine

Notes: Price is per-patient-price for a complete treatment. Number of patients cover total number of
patients that consume the drug in our sample. 1USD ≈ 65 INR (in 2017)

Table 2: Price Variation by Molecule

Average Price per patient (Indian Rupees)

Molecule Mean Std Min Max
Price Control from 1995
Pyrimethamine, Sulphadoxine 6 4 2 36
Chloroquine 9 2 3 16
Hydroxychloroquine 56 11 10 81

Not under price control

Quinine 201 38 91 351
A, S, P 210 34 123 360
Mefloquine 267 51 186 358
Arteether, Artemotil 154 45 38 390
Artesunate 198 47 75 398
Artemether, Lumefantrine 191 55 81 415
Artesunate, Mefloquine 457 127 293 615
Arteether, Lumefantrine 275 78 126 411
*A, S, P: Artesunate, Sulfadoxine, Pyrimethamine

Notes: Price is per-patient-price for a complete treatment. 1USD ≈ 65 INR (in 2017)

39
 Table 3: BLP Estimation Results

Parameters SE
Price*Rich Regions Dummy -8.69*** 2.33
Price*Poor Regions Dummy -24.58*** 1.88
No. of SKUs (injection) 7.50*** 1.12
No. of SKUs (tablet) 0.63 1.41
Log of Firm presence (in close therapies) 456.94*** 67.77
No of years since firm active 115.07*** 17.04
No of years since firm active squared -47.56*** 5.77
Demographic interactions
(with income)
Dummy for Price Control -2.20*** 0.48
Dummy New molecule in Rich region 0.67* 0.39
Dummy New molecule in Poor region 1.98*** 0.57
No of SKU 5.25*** 1.93
No of years since firm active -3.62*** 1.21
Region-Year Dummies Yes
Quarter Dummies Yes
Firm-Molecules Dummies Yes
Number of Observations 27,330
*** p<0.01, ** p<0.05, * p<0.1

Table 4: Own Price Elasticities in Rich and Poor Regions

Artesunate A-L A-A Ar-L

Tablet Injection Tablet Injection Tablet Injection Tablet Injection
Rich Region -2.16 -1.60 -1.63 -1.58 -1.31 -1.33 -1.65 -2.67
Poor Region -5.96 -4.50 -4.70 -4.42 -2.34 -3.67 -5.07 -7.59
Quinine Mefloquine A-M A-S-P
Tablet Injection Tablet Injection Tablet Injection Tablet Injection
Rich Region -1.67 -2.23 -2.35 - -3.33 -5.26 -1.62 -2.01
Poor Region -3.96 -6.12 -6.43 - -9.93 -14.79 -4.31 -5.67
*A-L: Artemether-Lumefantrine *A-A: Arteether-Artemotil *Ar-L: Arteether + Lumefantrine
*A-M: Artesunate + Mefloquine *A-S-P: Artesunate + Sulfadoxine + Pyrimethamine

Notes: This table presents mean own price elasticities wrt price in rich and poor region for the eight
molecules that are not under price control. For each molecule, we report own price elasticities for tablets
and injections separately. Since Mefloquine is offered only in tablet form, the elasticity numbers for
injection are left blank.

40
 Table 5: Average Mark-up by Molecule

Molecule Average Average Lerner’s

Price MC Ratio
Price Control Since 1995
Pyrimethamine, Sulphadoxine 6 5 8
Chloroquine 9 8 8
Hydroxychloroquine 56 51 8
Not Under Price Control
Quinine 201 139 34
A,S,P 210 137 36
Mefloquine 267 189 29
Artesunate 198 132 35
Artemether, Lumefantrine 191 124 37
Artesunate, Mefloquine 457 386 17
Arteether, Lumefantrine 275 206 27
Arteether, Artemotil 154 93 45
Number of Observations : 27330
*A,S,P: Artesunate, Sulfadoxine, Pyrimethamine
Notes: Lerner’s ratio: (price - mc)/price. This table presents average marginal costs and average markup
by molecule in Indian Rupees. 1 USD ' 65 INR (2017)

Table 6: Marginal Cost Parameter Estimates

Parameters SE
Dummy for Big firms 0.063*** 0.0061
No. of SKUs (injection) 0.199*** 0.0042
No. of SKUs (tablet) 0.043*** 0.0031
Log of Firm presence (in close therapies) -145.309*** 10.401
Number of years since firm active 0.012*** 0.0003
Number of years since firm active Squared -0.001*** 2e-05
R-Squared 0.90
Molecule dummies Yes
Quarter dummies Yes
Number of Observations 27,330

41
 Table 7: Results from Fixed Cost Estimation

Odisha Bihar Chhatisgarh

Constant -65.5 -21.6 -84.3 86.8 -7.5 46.3
No. SKU*Firm Presence 463.1 643.1 46.3 472.6 258.4 439.2
Years Since Brand active 69.5 94.3 16.9 52.6 39.2 57.4
AP Coastal UKhand-UPWest UP East
Constant -187.1 42.1 -70.2 -30.6 9.7 104.5
No. SKU*Firm Presence 91.9 1035.3 453 590.5 218.8 560.0
Years Since Brand active 16.5 75.8 19.4 36.7 -0.01 23.8
TamilNadu Punjab Madhya Pradesh
Constant -3.0 29.6 7.5 10.2 -176.7 89.0
No. SKU*Firm Presence 104.5 203.7 57.1 68.7 446.6 1290.9
Years Since Brand active -0.9 6.4 7.8 9.9 43.0 123.8
Kolkata Rajasthan Delhi - UP part
Constant -2.2 47 -111.7 146.9 97.9 104.2
No. SKU*Firm Presence 6.2 89.6 236.8 1133.3 -87.5 -5.2
Years Since Brand active -4.8 17 -14.2 74 18.4 26.2
Kerala Jharkhand West Bengal Rest
Constant -190.3 -17.0 -263.4 59.9 -28.0 62.0
No. SKU*Firm Presence 236.4 1215.7 266 1567.5 87.6 357.1
Years Since Brand active 1.2 10.0 -30.1 45.7 -17.8 21.6
Mumbai-Vicinity Haryana NorthEast
Constant -103.1 104.8 69.1 81.2 -4.5 0.8
No. SKU*Firm Presence 442.2 860.1 -95.3 -58.8 119 140.1
Years Since Brand active 72.3 134.8 35.1 38.5 7.9 9.8
Vidarbha AP Rest Marathwada
Constant -45.1 41.5 -90 -30.7 -46.4 60.8
No. SKU*Firm Presence 327.5 691.8 553.2 782.6 263.5 703.4
Years Since Brand active -2.7 86.5 62.7 77.1 33.2 81.1
Karnataka Gujarat
Constant 19.6 55.9 -64.2 -14.9
No. SKU*Firm Presence 95.6 162.0 171.2 477.8
Years Since Brand active -15.6 -7.4 -5.2 2.7
Sample Size: 83625
*Confidence intervals computed following, Kaido, Molinari, and Stoye (2016)
*Generalized moment selection applied following, Andrews and Soares (2010)

Notes: This table presents lower and upper parameter bounds from fixed cost estimation. We report
95% confidence intervals. For a given region, each row presents estimates of lower bound and upper
bound of the parameters.

42
 Table 8: Results from Fixed Cost Estimation

Avg FC Bounds Avg FC as % Fraction

(in ’000 INR) of Avg. Var profit
Odisha [ 209 240 ] [ 40.8 47.8 ]
Bihar [ 80 130 ] [ 45.9 74.6 ]
Chhatisgarh [ 169 190 ] [ 49.5 59.9 ]
AP Rest [ 183 204 ] [ 59.6 68.6 ]
AP Coastal [ 83 166 ] [ 16.3 39.3 ]
UKhand-UpWest [ 107 129 ] [ 37.2 46.2 ]
UP East [ 133 165 ] [ 38.2 55.9 ]
Marathwada [ 138 217 ] [ 59.2 90.3 ]
Kolkata [ 27 55 ] [ 23.2 47.0 ]
TamilNadu [ 31 66 ] [ 20.3 39.7 ]
Punjab [ 34 36 ] [ 31.4 33.0 ]
Madhya Pradesh [ 266 340 ] [ 46.1 75.1 ]
Jharkhand [ 68 157 ] [ 15.1 82.7 ]
Kerala [ 27 53 ] [ 28.0 53.9 ]
Rajasthan [ 191 276 ] [ 25.1 53.9 ]
Mumbai [ 267 320 ] [ 28.2 33.8 ]
Haryana [ 83 90 ] [ 41.2 44.2 ]
NorthEast [ 45 50 ] [ 45.5 50.2 ]
Delhi [ 114 123 ] [ 61.2 65.9 ]
Vidarbha [ 121 229 ] [ 44.6 84.5 ]
Karnataka [ 42 62 ] [ 18.3 27.2 ]
Gujarat [ 24 71 ] [ 14.1 39.3 ]
Notes: First column in this table presents average per product fixed cost bounds in thousands Indian
Rupees in a given region per quarter. The second column reports per product fixed cost as a fraction of
average per product variable profit.

43
 Table 9: Product Offering for Different Molecules across Price Margin Values in Odisha

1% 2% 3% 4% 5%
Artesunate [ 0 0 ] [ 0 0 ] [ 0 1 ] [ 1 1 ] [ 1 1 ]
Artemether-Lumefantrine [ 0 0 ] [ 0 0 ] [ 0 0 ] [ 0 0 ] [ 0 0 ]
Arteether-Artemotil [ 0 0 ] [ 0 0 ] [ 0 0 ] [ 0 1 ] [ 0 1 ]
6% 7% 8% 9% 10%
Artesunate [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ]
Artemether-Lumefantrine [ 0 0 ] [ 0 0 ] [ 0 0 ] [ 0 0 ] [ 0 1 ]
Arteether-Artemotil [ 0 1 ] [ 1 2 ] [ 1 2 ] [ 1 2 ] [ 1 3 ]
11% 12% 13% 14% 15%
Artesunate [ 1 2 ] [ 1 2 ] [ 1 2 ] [ 1 3 ] [ 1 3 ]
Artemether-Lumefantrine [ 0 1 ] [ 0 1 ] [ 0 1 ] [ 0 1 ] [ 1 1 ]
Arteether-Artemotil [ 1 3 ] [ 1 3 ] [ 1 3 ] [ 1 3 ] [ 1 3 ]
16% 17% 18% 19% 20%
Artesunate [ 1 3 ] [ 1 3 ] [ 1 3 ] [ 2 3 ] [ 2 3 ]
Artemether-Lumefantrine [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ]
Arteether-Artemotil [ 1 3 ] [ 1 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ]
21% 22% 23% 24% 25%
Artesunate [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ]
Artemether-Lumefantrine [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ]
Arteether-Artemotil [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ]
26% 27% 28% 29% 30%
Artesunate [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ]
Artemether-Lumefantrine [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ] [ 1 1 ]
Arteether-Artemotil [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ] [ 2 3 ]
Notes: This table reports bounds of number of products offered across equilibria at different margin values
for three molecules separately. Each entry in the table refers to the minimum and maximum number of
products for a specific molecule at a given margin value. For example the last entry [2 3] implies that
at 30% margin, for Arteether-Artemotil, minimum number of products offered across equilibria is 2 and
maximum number of products offered across equilibria is 3.

References
Andrews, D. W., S. Berry, and P. Jia (2004): “Confidence regions for parameters in discrete games
with multiple equilibria, with an application to discount chain store location,” .

Andrews, D. W., and G. Soares (2010): “Inference for Parameters Defined by Moment Inequalities
Using Generalized Moment Selection,” Econometrica, 78(1), 119–157.

44
Antony, H. A., and S. C. Parija (2016): “Antimalarial drug resistance: An overview,” Tropical Para-
sitology, 6(1), 30.

Bajari, P., H. Hong, and S. P. Ryan (2010): “Identification and estimation of a discrete game of
complete information,” Econometrica, 78(5), 1529–1568.

Basak, S., and D. Sathyanarayana (2012): “Exploring knowledge and perceptions of generic medicines
among drug retailers and community pharmacists,” Indian journal of pharmaceutical sciences, 74(6), 571.

Beresteanu, A., I. Molchanov, and F. Molinari (2011): “Sharp identification regions in models with
convex moment predictions,” Econometrica, 79(6), 1785–1821.

Berndt, E. R., and I. M. Cockburn (2014): “The Hidden Cost Of Low Prices: Limited Access To New
Drugs In India,” Health Affairs, 33(9), 1567–1575.

Berry, S., J. Levinsohn, and A. Pakes (1995): “Automobile prices in market equilibrium,” Economet-
rica: Journal of the Econometric Society, pp. 841–890.

Bhaskarabhatla, A., C. Chatterjee, and B. Karreman (2016): “Hit Where It Hurts: Cartel Policing
Using Targeted Sales and Supply Embargoes,” The Journal of Law and Economics, 59(4), 805–846.

Bond, E. W., and K. Saggi (2014): “Compulsory licensing, price controls, and access to patented foreign
products,” Journal of Development Economics, 109, 217–228.

Boni, M. F., D. L. Smith, and R. Laxminarayan (2008): “Benefits of using multiple first-line therapies
against malaria,” Proceedings of the National Academy of Sciences, 105(37), 14216–14221.

Chatterjee, C., K. Kubo, and V. Pingali (2015): “The Consumer Welfare Implications of Govern-
mental Policiesand Firm Strategy in Markets for Medicines,” Journal of Health Economics.

Chaudhuri, S., P. K. Goldberg, and P. Jia (2006): “Estimating the Effects of Global Patent Protection
in Pharmaceuticals: A Case Study of Quinolones in India,” The American Economic Review.

Ciliberto, F., and E. Tamer (2009): “Market structure and multiple equilibria in airline markets,”
Econometrica, 77(6), 1791–1828.

Cockburn, I. M., J. O. Lanjouw, and M. Schankerman (2016): “Patents and the global diffusion of
new drugs,” American Economic Review, 106(1), 136–64.

Conlon, C. T. (2013): “The empirical likelihood mpec approach to demand estimation,” .

Crawford, G. S. (2012): “Endogenous product choice: A progress report,” International Journal of

Industrial Organization, 30(3), 315–320.

Danzon, P. M., A. W. Mulcahy, and A. K. Towse (2015): “Pharmaceutical pricing in emerging

markets: effects of income, competition, and procurement,” Health economics, 24(2), 238–252.

Danzon, P. M., Y. R. Wang, and L. Wang (2005): “The impact of price regulation on the launch delay
of new drugs—evidence from twenty-five major markets in the 1990s,” Health economics, 14(3), 269–292.

45
DellaVigna, S., and M. Gentzkow (2017): “Uniform pricing in us retail chains,” Discussion paper,
National Bureau of Economic Research.

Dhingra, N., P. Jha, V. P. Sharma, A. A. Cohen, R. M. Jotkar, P. S. Rodriguez, D. G. Bassani,

W. Suraweera, R. Laxminarayan, R. Peto, et al. (2010): “Adult and child malaria mortality in
India: a nationally representative mortality survey,” The Lancet, 376(9754), 1768–1774.

Dickstein, M. J., and E. Morales (2013): “Accounting for expectational and structural error in bi-
nary choice problems: A moment inequality approach,” Discussion paper, National Bureau of Economic
Research.

(2015): “What do exporters know?,” .

Draganska, M., M. Mazzeo, and K. Seim (2009): “Beyond plain vanilla: Modeling joint product
assortment and pricing decisions,” QME, 7(2), 105–146.

Dubois, P., O. De Mouzon, F. Scott-Morton, and P. Seabright (2015): “Market size and phar-
maceutical innovation,” The RAND Journal of Economics, 46(4), 844–871.

Dubois, P., and M. Saethre (2018): “On the Effect of Parallel Trade on Manufacturers’ and Retailers’
Profits in the Pharmaceutical Sector,” .

Duggan, M., C. Garthwaite, and A. Goyal (2016): “The Market Impacts of Pharmaceutical Product
Patents in Developing Countries: Evidence from India,” American Economic Review, 106(1), 99–135.

Dutta, A. (2011): “From free entry to patent protection: Welfare implications for the Indian pharmaceutical
industry,” The Review of Economics and Statistics, 93(1), 160–178.

Eizenberg, A. (2014): “Upstream Innovation and Product Variety in the US Home PC Market*,” The
Review of Economic Studies, p. rdu004.

Fan, Y. (2013): “Ownership consolidation and product characteristics: A study of the US daily newspaper
market,” The American Economic Review, 103(5), 1598–1628.

Fan, Y., and C. Yang (2016): “Competition, product proliferation and welfare: A study of the us
smartphone market,” .

Filson, D. (2012): “A Markov-perfect equilibrium model of the impacts of price controls on the performance
of the pharmaceutical industry,” The RAND Journal of Economics, 43(1), 110–138.

Goldberg, P. K. (2010): “Alfred Marshall Lecture intellectual property rights protection in developing
countries: the case of pharmaceuticals,” Journal of the European Economic Association, 8(2-3), 326–353.

Gupta, I., S. Chowdhury, et al. (2014): “Economic burden of malaria in India: the need for effective
spending,” WHO South-East Asia journal of public health, 3(1), 95.

Ho, K., and A. Pakes (2014): “Hospital Choices, Hospital Prices, and Financial Incentives to Physicians,”
The American Economic Review, pp. 3841–3884.

46
Kaido, H., F. Molinari, and J. Stoye (2016): “Confidence intervals for projections of partially identified
parameters,” arXiv preprint arXiv:1601.00934.

Kaido, H., F. Molinari, J. Stoye, and M. Thirkettle (2017): “Calibrated Projection in MATLAB:
Users’ Manual,” arXiv preprint arXiv:1710.09707.

Kessler, D. P. (2004): “The effects of pharmaceutical price controls on the cost and quality of medical
care: a review of the empirical literature,” .

Kremer, M. (1998): “Patent buyouts: A mechanism for encouraging innovation,” The Quarterly Journal
of Economics, 113(4), 1137–1167.

(2002): “Pharmaceuticals and the developing world,” Journal of Economic Perspectives, 16(4),
67–90.

Kumar, A., N. Valecha, T. Jain, and A. P. Dash (2007): “Burden of malaria in India: retrospective
and prospective view,” The American journal of tropical medicine and hygiene, 77(6 Suppl), 69–78.

Kyle, M., and Y. Qian (2014): “Intellectual Property Rights and Access to Innovation: Evidence from
TRIPS,” (20799).

Kyle, M. K. (2007): “Pharmaceutical price controls and entry strategies,” The Review of Economics and
Statistics, 89(1), 88–99.

Kyle, M. K., and A. M. McGahan (2012): “Investments in pharmaceuticals before and after TRIPS,”
Review of Economics and Statistics, 94(4), 1157–1172.

Lanjouw, J. O. (1998): “The Introduction of Pharmaceutical Product Patents in India:” Heartless Ex-
ploitation of the Poor and Suffering”?,” .

(2005): “Patents, price controls and access to new drugs: how policy affects global market entry,” .

Manski, C. F., and J. V. Pepper (2000): “Monotone instrumental variables: with an application to the
returns to schooling,” Econometrica, 68(4), 997–1010.

(2009): “More on monotone instrumental variables,” The Econometrics Journal, 12(s1).

Mazzeo, M. J. (2002): “Product choice and oligopoly market structure,” RAND Journal of Economics,
pp. 221–242.

Mehta, A., H. H. Farooqui, and S. Selvaraj (2016): “A critical analysis of concentration and compe-
tition in the Indian pharmaceutical market,” PloS one, 11(2), e0148951.

Morton, F. M. S. (1999): “Entry decisions in the generic pharmaceutical industry,” The Rand journal of
economics, pp. 421–440.

Moser, P. (2013): “Patents and innovation: evidence from economic history,” Journal of Economic Per-
spectives, 27(1), 23–44.

47
Nevo, A. (2000a): “Mergers with differentiated products: The case of the ready-to-eat cereal industry,”
The RAND Journal of Economics, pp. 395–421.

(2000b): “A Practitioner’s Guide to Estimation of Random-Coefficients Logit Models of Demand,”

Journal of Economics & Management Strategy, 9(4), 513–548.

(2001): “Measuring market power in the ready-to-eat cereal industry,” Econometrica, 69(2), 307–
342.

Nosko, C. (2011): “Competition and quality choice in the cpu market,” .

Organization, W. H., et al. (1990): WHO model prescribing information: drugs used in parasitic
diseases. World Health Organization.

Organization, W. H., UNICEF, et al. (2015): Achieving the malaria MDG target: reversing the inci-
dence of malaria 2000-2015. World Health Organization.

Pakes, A. (2010): “Alternative models for moment inequalities,” Econometrica, 78(6), 1783–1822.

Pakes, A., J. Porter, K. Ho, and J. Ishii (2015): “Moment Inequalities and Their Application,”
Econometrica, 83(1), 315–334.

Petersen, I., R. Eastman, and M. Lanzer (2011): “Drug-resistant malaria: molecular mechanisms
and implications for public health,” FEBS letters, 585(11), 1551–1562.

Seim, K. (2006): “An empirical model of firm entry with endogenous product-type choices,” The RAND
Journal of Economics, 37(3), 619–640.

Shrank, W. H., E. R. Cox, M. A. Fischer, J. Mehta, and N. K. Choudhry (2009): “Patients’

perceptions of generic medications,” Health Affairs, 28(2), 546–556.

Small, K. A., and H. S. Rosen (1981): “Applied Welfare Economics with Discrete Choice Models,”
Econometrica, 49(1), 105–30.

Tamer, E. (2003): “Incomplete simultaneous discrete response model with multiple equilibria,” The Review
of Economic Studies, 70(1), 147–165.

Wollmann, T. (2017): “Trucks without bailouts: Equilibrium product characteristics for commercial
vehicles,” American Economic Review (Forthcoming).

48
A Appendix: For On-line Publication Only
A.1 Data Preparation Details
In a given market, pharmaceutical products are available in multiple presentations (SKUs),
that is, combination of dosage forms (e.g. tablets and injections), strength ( e.g. 100
milligrams, 500 milligrams) and packet sizes (e.g. 50 tablet bottle, 100 tablet bottle). Our
data contain information on firm-level sales as well as maximum retail prices (MRP) of these
molecules at the stock keeping unit (SKU) level. Since we define a product at the level of firm-
molecule-tablet and firm-molecule-injection, in our analysis, we aggregate SKUs for tablets
and for injections and use this for our estimation. For example, if IPCA pharmaceuticals
produces 10 pack and 20 pack tablets for chloroquine, we combine these SKUs for tablets to
define a product as IPCA-Chloroquine-Tablet.
To get the product level variables, we combine the sales and prices across the tablet and
injection SKUs in a market for a given molecule produced by a firm. To convert the different
presentations into number of dosages, we consult pharmacological literature and assign a
patient-day dosage measure to each product in the sample. For example, if one prescribed
dosage involves two tablets each of 100 mg strength, then a 10 tablet pack would have 5
dosages in total. We also collect the information on the number of dosages for a complete
treatment of a patient for each molecule. We refer to WHO model prescribing information
(Organization et al. (1990)) and Hospital of the University of Pennsylvania Malaria Adult
Treatment Guidelines58 for collecting this information. For example, 2500 mg of chloroquine
over 3 days need to be administered per patient for a complete treatment with chloroquine.
We convert the sales data for each SKU into number of patients and add across solid and
liquid SKUs to get total sales for products defined as firm-molecule-tablet and firm-molecule-
injection. Similarly, we obtain price per patient for each SKU by dividing price of each SKU
with the number of complete dosages for the given SKU. We then obtain average price across
tablets as well as injections and define that as the price of the product.
To obtain market size, we collect number of malarial patients reported across different
regions in India from the Indian National Vector Borne Disease Control website. In some
of the markets, number of malarial patients obtained from this source are less as compared
to the total sales (in number of malarial patients) that we observed in our data. This
discrepancy might be due to under-reporting of malarial cases in the government data. To
handle this, and to make sure that the outside option has a non-negative share, we used a
multiplier to inflate the number of malarial patients and used this as a measure of market
58
http://www.uphs.upenn.edu/bugdrug/antibiotic_manual/malariaRx-HUP.htm

49
size. Note that, this approach keeps the ratio of malarial patients unchanged across regions.
In our empirical application, we fix the multiplier to be equal to 4. We experimented with
different multiplier values and our demand estimates are robust to these specifications.

A.2 Details of Marginal Cost Estimation

In this section, we discuss the details of marginal cost estimation as described in section
4.2. Note that, a key challenge we face is that, for the products not under price control, a
firm charges identical price across all regions in India where it chooses to offer the product.
We discuss the details of computation of mark-up given demand estimates, that we obtain
from a modified set of first order conditions. Once mark-ups are estimated, marginal cost is
obtained from the observed prices as a residual.
For a given firm f , the profit maximization problem is given by equation 3.2. For a
product under molecule g where g is not under price control, firm f sets price pf gt across all
regions where firm f operates (Rf t ) at time period t to maximize profit. Corresponding first
order condition is given by:

∂πf t X ∂sf grt X

= (pf gt − mcf gt ) Mrt + sf grt Mrt
∂pf gt r∈R ∂pf gt r∈R
ft ft
  (A.1)
X X ∂sf krt X X ∂sf lrt pf lt ∗ marg%
+ Mrt (pf kt − mcf kt ) + Mrt =0
r∈R k∈J nc
∂p f gt
r∈R l∈J c
∂p f gt 1 + marg%
ft f rt ft f rt

In the above expression, Rf t denotes the set of regions where firm f operates at time t. Price,
marginal cost and market share of molecule g offered by firm f at time t are denoted by
pf gt , mcf gt , sf rgt respectively. Jfcrt and Jfncrt denote the set of molecules offered by firm f in
region r and time t that are under price control and not under price control respectively. Mrt
denotes the market size of region r at time t. We denote the set of products offered by all
firms in India across all regions that are not under price control and under price control as
Jnc (≡ ∪f,r,t Jfncrt ) and Jc (≡ ∪f,r,t Jfcrt ) respectively. Further, we denote by Jnc c
u and Ju the set of
firm-molecule-time specific unique products across regions that are not under price control
and under price control respectively. For example, if firm 1 offers molecule a in region r1
and in region r2 , and molecule b in region r1 that are not unde price control, then Jnc would
include (f1 ar1 , f1 ar2 , f1 br1 ), while Jnc nc
u would include (f1 a, f1 b). Our purpose of defining Ju
is to make the point that since prices are identical across regions, for each product included
in Jncu , we have a corresponding price data in sample.
We denote the (|Jnc | × |Jnc |)  dimensional  derivative matrix by ∆ where the (i, j)th el-
∂sf 0 jrt
ement of the matrix is given by ∂pf it Mrt that denotes the partial derivative of market

50
share of product j wrt price of product i multiplied by corresponding market size. We denote
by T , (|Jnc | × |Jnc |) dimensional ownership matrix, where Ti,j = 1, if i and j are offered by
the same firm and Ti,j = 0, if i and j are offered by different firms.
Let us denote by vector S, with dimension (|Jnc u | × 1). Suppose, the k-th element in this
vector corresponds to a molecule g and belongs to firm f , and time period t. Then the k-th
element is given by

X X X ∂scf lrt 
pf lt ∗ marg%

− snc
f grt Mrt − Mrt .
r∈Rf t r∈R l∈J c
∂pf lt 1 + marg%
ft f rt

In the above expression, sncf grt denotes the market share of molecule g, offered by firm f in
region r at time t where g is not under price control. Similarly, scf lrt denotes the market share
of product l offered by firm f , in region r at time t, where l ∈ Jfcrt that is, l is under price
control. Let p~ be the vector of prices for all the products that are not under price control,
and mc~ be the corresponding marginal cost for the products. Note that both p~ and mc ~ are of
dimension (|Ju | × 1). Finally, we define D, an indicator matrix of dimension ((|J | × |Jnc
nc nc
u |),
where for a given column j, all the rows that correspond to the jth unique product, that is
all the products that are offered by a given firm from a specific molecule at a given point in
time (across different regions) are denoted 1, and 0 otherwise. Putting everything in matrix
form, we have,
~ = DT (T ∗ ∆)D S


(~p − mc) (A.2)

where ∗ represents element-by-element multiplication.

A.3 Selection problem and empty intervals - example

If we are willing to ignore the selection problem and just assume the conditional expecta-
tions to be equal to 0 as in equation 4.7, and use the corresponding moment conditions for
estimation of parameters, we may end up with empty intervals for fixed cost. To see this
in our context, let us consider a specific example. With some abuse of notations, let us
suppose that in a sample, the econometrician’s estimate of E(∆πu\ |dj = 1, If t ) = 500 and
E(∆πl\ |dj = 0, If t ) = 550. Suppose the fixed cost to be estimated (W θ) and unknown to the
econometrician is equal to 525. Additionally, let us assign values to the conditional expecta-
tions of the error terms given by [E(νf jrt |dj = 1, If t ) = −30] and [E(νf jrt |dj = 0, If t ) = 30].
Note that the specific numbers we have picked in this example satisfies all the conditions
in equations 4.5 and 4.8. However, if we assume conditional expectations equal to 0 in our

51
example, then while estimating, we end up with moment conditions given by

W θ ≤ 500 and W θ ≥ 550; leading to empty interval

even when the actual interval,
W θ ≤ 500 − (−30) and W θ ≥ 550 − (30); is nonempty

A.4 Inequality Estimation Details

A.4.1 Computation of Profit Bounds

Here we describe the computation of differences in expected variable profit from adding and
dropping a product. Note that, we denote by by Jf,rt , the set of products offered by firm f
in market rt and by Jfp , the potential set of products by firm f . Further, at time period
t, a firm operates across different regions and we denote this set by Rf t . For a product in
the menu of potential products, and (not) offered in a market, we compute the difference in
expected variable profit from (adding) dropping the product. We discuss the details for a
product offered in the market, for a product not offered in the market, computational details
are similar. Hence, for each j ∈ Jf,rt , we need to compute the drop in expected variable
profit from not offering the product in the market given by:
h
 i
− π Jf,rt \ j, {Jf,st }s∈Rf t \r , {J−f,st }s∈Rf t If t


Eξ,ω π {Jf,st }s∈Rf t , {J−f,st }s∈Rf t

For each offered product, we need to first compute the equilibrium price and market share


vector under offered product portfolio is given by {Jf,st }s∈Rf t , {J−f,st }s∈Rf t , that is the equi-
librium prices that firm f and its competitor’s charge across Rf t regions, and corresponding
market shares. Similarly, we also need to compute equilibrium prices and shares, for the port-


folio after dropping a product in a specific market given by Jf,rt \ j, {Jf,st }s∈Rf t \r , {J−f,st }s∈Rf t .
Note that, we define our market as a region-quarter (rt) pair. However, since firms charge
same price across all regions in a given quarter, firms solve a joint maximization problem
across Rf t regions. We use the first order conditions from the profit maximization problem
to compute equilibrium prices. Note that from equation A.2, we have

DT (T ∗ ∆)D S p~, θd , γ



(~p − mc)
~ =
| {z }
mark-up

~ estimates, as well as estimates from demand (θd , γ), we use contraction mapping
Given mc
to compute equilibrium price vector. We start with the observed price vector (in the data),
compute the markup, and given markup compute the new price vector by using A.2. We

52
repeat this until the norm of difference between successive price vectors is less than toler-
ance level (10−6 ). We follow similar steps to compute the equilibrium price vector after we
drop a product j from a specific market. Since we compute equilibrium price vector across
multiple markets solving first order conditions from a joint maximization problem, while
simulation, we assume demand residuals (ξ) and supply residual (ω) to be equal to zero.
The empirical distribution of demand and supply residuals from our model estimates are
very tightly distributed centered around 0. Hence, this approximation is primarily done to
alleviate computational burden. Another option is to draw ns vector of demand and supply
ˆ ω̂), compute equilibrium profit from
residuals from the estimated empirical distribution of (ξ,
one step deviation for each error vector, and take the average across ns vectors to compute
the expected profit. This is a computationally feasible option when the computation can
be done for each market separately. In our case, due to computation of equilibrium price
and market share across multiple regions, this process becomes practically infeasible due to
computational reasons. Once the equilibrium price and market shares are computed, we get
firm specific profit both before and after dropping the product. The difference gives us the
change in expected variable profit for the specific product. We repeat these steps for each of
the products that are included in the menu of potential products across all firms.

A.4.2 Conditional Moment Inequality for Lower Bound

For lower bound, the MIV assumption similar to 4.10 is given by

MIV assumption (Lower Bound): Let V be an ordered set with strictly positive sup-
port. Covariate v is a monotone instrumental variable in the sense of mean-monotonicity,
if for all (v1 , v2 ) ∈ (V × V ), with 0 < v1 ≤ v2 ,

E(νf jrt dj = 0, If t , v = v1 ) ≤ E(νf0 jrt dj = 1, If t , v = v2 )

     
(A.3)

where |.| denotes the absolute value function.

Using A.3 and the normalization assumption in 4.11, we derive the inequalities for estimation
of lower bound of fixed cost parameters. Note that, given A.3, with 0 < v1 ≤ v2 , we have

E(νf jrt |dj = 0, If t , v = v1 ) ≤ −E(νf0 jrt |dj = 1, If t , v = v2 )

=⇒ E(νf jrt v1 |dj = 0, If t , v = v1 ) ≤ −E(νf0 jrt v2 |dj = 1, If t , v = v2 ) (A.4)
=⇒ E(νf jrt v1 |dj = 0, If t , v = v1 ) + E(νf0 jrt v2 |dj = 1, If t , v = v2 ) ≤ 0

53
 Using the expression for lower bound in 4.5, and equation 4.11, we have,

E(∆πl − Wf jrt θIf t , v1 ) − E(νf jrt dj = 0, If t , v = v1 ) ≤ 0, and

 

E(∆πjpc − Wfpcjrt θIf t , v2 ) − E(νf jrt |dj = 1, If t , v = v2 ) = 0



=⇒ E(∆πl v1 − Wf jrt θv1 If t , v1 ) − E(νf jrt v1 |dj = 0, If t , v = v1 ) +

  
(A.5)
E(∆πjpc v2 − Wfpcjrt θv2 If t , v2 ) − E(νf jrt v2 |dj = 1, If t , v = v2 ) ≤ 0
  

=⇒ E (∆πl v1 + ∆πjpc v2 ) − (Wf jrt v1 + Wfpcjrt v2 )θIf t , v1 , v2 ≤

  

[E(νf jrt v1 |dj = 0, If t , v = v1 ) + E(νf jrt v2 |dj = 1, If t , v = v2 )] ≤ 0

Given A.4, the inequalities for lower bound are given by

E (∆πl v1 + ∆πjpc v2 ) − (Wf jrt v1 + Wfpcjrt v2 )θIf t , v1 , v2 ≤ 0

  
(A.6)

A.4.3 Derivation of Unconditional Moments

The moment inequalities described in equations 4.15, and 4.16 condition on particular values
of the instrument vector v. We follow the details explained in Dickstein and Morales (2015)59 ,
and derive unconditional moment inequalities. Each of the unconditional moments is defined
by an instrument function. Specifically, given a positive-valued instrument function g(.), we
derive unconditional moments that are consistent with our unconditional moments:

mu (Wf jrt , θ; If t )
"( ) #
E × g(vf jrt ) ≤ 0
ml (Wf jrt , θ; If t )

where mu (.), ml (.) are defined as in equations 4.15, and 4.16, and vf jrt includes the set of
monotone instruments. For each instrument vector vk,f jrt included in the instrument vector
vf jrt
1{vk,f jrt > med(vk,f jrt }
( )
 
a
ga (vk,f jrt ) = × vk,f jrt − med(vk,f jrt ) .
1{vk,f jrt ≤ med(vk,f jrt }
In words, for each scalar random variable vk,f jrt included in the instrument set vf jrt , the
function ga (.) builds two moments by splitting the observations into two groups depending
on whether the value of the instrument variable for the observation is above or below its
median. We use different instrument functions ga (vf jrt ), for a = {0, 1}.
59
See their appendix A.4

54
A.4.4 Computation of Confidence Set

Note that our moment conditions are written in the form

E(m(Wf jrt , θ)) ≤ 0

The identified set is ΘI ={θ : E(m(Wf jrt , θ)) ≤ 0}

where Wf jrt are the covariates used in our fixed cost estimation, and m = (m1 , . . . , mK ) are
the K moment conditions we use to estimate the bounds for θ. In our analysis, we allow
for Fixed cost parameters to vary across regions. Our specification for Ff jrt is given by 3.4.
Given instruments, we construct K sample moment conditions given by

1 X
m̄k = mk (Wf jrt , θ, If t ); k = 1, . . . , K
n f,j,r,t

where n denotes the sample size. Since our fixed cost parameters are region specific, and for
each region, there are 3 parameters to be estimated, we end up estimating 69 parameters
for 23 regions in our sample. We use the techniques developed in Kaido, Molinari, and
Stoye (2016) for computation of confidence intervals. This procedure implements a support-
function based method for computation of confidence intervals using the projection of the
identified set. We report element wise confidence interval for our parameter vector θ. To do
this, we consider the projection of ΘI in a specific direction. For example, for first element
of θ, we are interested in the projection of ΘI in the direction p and −p, where p is the unit
vector with first element equal to 1. As confidence interval for this object, we report the
projection of a relaxation of the sample inequality conditions, that is for the k-th element of
θ, we solve:

max / min θk
θ ∈ ΘI
√
nm̄j (x, θ)
s. t. ≤ ĉn (θ)
σ̂j (θ)

where σ̂j (θ) is an estimator of asymptotic standard deviation of the moment conditions, and
is given by
1X
σ̂j (θ) = (mk (Wf jrt , θ) − m̄(θ))) (mk (Wf jrt , θ) − m̄(θ)))0
n f jrt

Here ĉn (θ) is the critical value and the 95% coverage is achieved by properly calibrating ĉn (θ).
The critical value is computed by checking feasibility of a linear program across bootstrap

55
repetitions. While calibrating the critical value, we select the binding moments following
the generalized moment selection as discussed in Andrews and Soares (2010). To alleviate
the computational burden, we use the EAM algorithm suggested in Kaido, Molinari, and
Stoye (2016) while implementing the computation of confidence interval. The details of
the algorithm and computational specifics are documented in Kaido, Molinari, Stoye, and
Thirkettle (2017).

56